Pith. sign in

REVIEW 5 major objections 4 minor 15 references

Graph Coloring for Multi-Task Learning

T0 review · 5 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A scheduler that activates only one low-conflict color class of tasks per training step improves multi-task learning and preserves the standard SGD convergence rate.

desk verdict Neat scheduler idea and a solid recovery lemma, but the descent and convergence proofs do not cover the algorithm as run—the claimed guarantees are the paper's main soft spot. read the letter →

arxiv 2509.16959 v5 pith:HOXTUU35 submitted 2025-09-21 cs.LG cs.AIcs.NEstat.ML

Santosh Patapati , Ian Noronha This is my paper · ORCID
classification cs.LGcs.AIcs.NEstat.ML
keywords multi-tasklearningnegativetransfergraphcoloringgradientinterferencetaskschedulingconflictWelsh-PowellalgorithmSGDconvergence
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces SON-GOKU, a scheduler that measures pairwise gradient interference between tasks, builds a conflict graph whose edges join tasks whose averaged gradients point against each other, greedily colors that graph, and updates exactly one color class per training step. Its central claim is that this sequential, low-conflict scheduling reduces negative transfer and improves the final performance of any underlying multi-task optimizer, without retuning the optimizer. The paper backs this with theory: grouped low-conflict updates preserve descent, the method reaches the usual nonconvex SGD rate $O(1/\sqrt{T})$ up to a small conflict-dependent constant, and over a refresh window scheduled group updates are never worse than a single mixed update and are strictly better when between-group interference is negative. Empirically, the authors report consistent gains over baselines and state-of-the-art multi-task optimizers across six datasets.

What carries the argument

The load-bearing object is the conflict graph $G_\tau$: vertices are tasks, and an edge exists between tasks $i,j$ when the interference coefficient $\rho_{ij} = -\langle \tilde{g}_i, \tilde{g}_j\rangle / (\|\tilde{g}_i\|\|\tilde{g}_j\|)$ exceeds the threshold $\tau$, where $\tilde{g}_i$ are EMA-smoothed shared-parameter gradients. Greedy Welsh-Powell coloring partitions the tasks into color classes with no conflict edges inside a class, and the scheduler activates one color class per step in a cyclic order, refreshing the graph every $R$ steps. This mechanism converts the continuous problem of choosing which tasks should update together into a discrete combinatorial one, with the $\Delta+1$ color bound providing both a bounded update delay and a compact schedule that scales with the worst conflict degree rather than the number of tasks.

What would settle it

Construct two tasks whose mean gradient directions are aligned (cosine above $-\tau$) but whose Hessian cross-terms are strongly negative, so a joint update increases the loss; run SON-GOKU with both in one color class and check whether the descent bound of Equation (12) is violated. Alternatively, fix a refresh window of length $R$ and measure whether task gradient directions rotate by more than the recovery margin $\gamma$; if they do, the exact-recovery guarantee predicts that the empirical conflict graph should diverge from the population graph.

Watch

Extended reading notes

Core claim

The discovery is that conflict-aware scheduling can be reduced to graph coloring: compute EMA-smoothed gradient cosines, threshold them at a tolerance $\tau$ to define a conflict graph, apply the Welsh-Powell largest-first greedy coloring, and cyclically activate one color class per step while periodically rebuilding the graph. The paper proves that a $\tau$-compatible group's aggregate update is a guaranteed descent direction, with the squared norm of the joint update bounded below by $(1-\tau(|S|-1))$ times the sum of per-task squared norms. With step size $\eta = c/\sqrt{T}$, convergence follows the standard nonconvex SGD rate with a mild $(1+\tau)$ factor, and under a separation margin between conflicting and non-conflicting pairs the empirical conflict graph exactly recovers the population graph with high probability. The paper further claims that sequential group updates dominate a single aggregated update in expected descent when cross-group interactions are sufficiently negative, and that the cyclic schedule keeps every task updated at least once every $\Delta+1$ steps, where $\Delta$ is the maximum conflict degree.

Load-bearing premise

The load-bearing premise is that the angle between recent averaged task gradients, thresholded at a chosen tolerance, faithfully measures harmful interference and stays stable within each refresh window; if real conflicts live in higher-order interactions or the gradients swing inside a window, the coloring is built on the wrong graph.

Editorial extensions

If this is right

  • SON-GOKU behaves as a plug-in scheduler: it can wrap existing multi-task optimizers such as PCGrad, AdaTask, and GradNorm, and the paper reports that the combinations improve over the base methods without additional tuning.
  • The convergence guarantee means the scheduler does not sacrifice the asymptotic $O(1/\sqrt{T})$ rate; interference control shows up only in the constant via the $(1+\tau)$ factor, so practitioners can trade a small constant for reduced negative transfer.
  • The $\Delta+1$ cyclic schedule provides a bounded staleness guarantee: every task is updated at least once per period, so the method is compatible with analyses that assume bounded inter-update delay.
  • The exact-recovery result gives a concrete probe budget: if the EMA effective sample size satisfies $n_{\text{eff}} \gtrsim \sigma^2/(m_0^2\gamma^2)\log(K/\delta)$, the estimated conflict graph matches the true one with high probability, linking a practical hyperparameter choice to measurable noise, margin, and task-count quantities.
  • The ablation study shows that dynamic recoloring and history-smoothed conflict estimates are load-bearing: freezing the coloring or using only the most recent mini-batch gradients degrades performance, supporting the paper's claim that task relationships drift and need periodic refresh.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The theory implicitly predicts that the benefit of scheduling grows with the strength of negative cross-group interference; a direct test would be to construct synthetic task sets with controlled pairwise cosine angles and verify that the advantage over a single mixed update tracks the predicted descent gap.
  • Because the schedule length depends on the maximum conflict degree $\Delta$ rather than $K$, the method should scale gracefully to many tasks when most pairs are non-conflicting; measuring the empirical $\Delta$ on real workloads would test this scaling premise.
  • The cosine proxy itself is the main extension risk: tasks with aligned mean gradients can still interfere through higher-order Hessian interactions, so a natural next step would be to replace flat gradient cosines with layer-wise or Hessian-weighted conflict measures and check whether the coloring changes materially.
  • The recovery bound suggests that the EMA refresh window and the tolerance $\tau$ should be tuned together with the expected drift of task gradients; one testable implication is that an adaptive refresh period based on measured gradient drift would preserve the guarantees at lower average cost.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper proposes SON-GOKU, a scheduler for multi-task learning that estimates pairwise gradient interference via EMA-smoothed cosine similarity, builds a thresholded conflict graph, greedily colors it, and activates one color class per training step, with periodic recoloring. It claims theoretical guarantees that τ-compatible active sets preserve descent, that a nonconvex O(1/√T) rate holds with a mild (1+τ) constant, that scheduled group updates are never worse and often better than a single mixed update, that the population conflict graph and task partition can be recovered exactly under a margin assumption, and that greedy coloring gives bounded staleness. Empirically, it reports consistent improvements over baselines across six datasets and further gains when combined with PCGrad, AdaTask, and GradNorm.

Significance. If the theoretical claims held, the paper would offer a simple, low-overhead plug-in scheduler with a rare theoretical treatment of dynamic task grouping, and the six-dataset empirical study with ablations would be a useful contribution. The graph-recovery analysis in Appendix B is internally coherent under its explicit assumptions, and the ablation results support the importance of dynamic recoloring and EMA smoothing. However, several load-bearing theoretical statements are not established as written: the descent claim conflates a norm bound with a full-objective inner-product condition, a key refinement inequality (Proposition 7) is false, the convergence rate in Eq. (13) is not derived for the deterministic schedule actually implemented, and the scheduled-versus-aggregated comparison in Section 5.3 is contradicted by the paper's own Appendix C.4. These issues are central to the paper's claim to explain why grouping and sequential updates improve multi-task learning.

major comments (5)
  1. [Appendix E, Proposition 7] Eq. (12) lower-bounds the squared norm of the active-gradient sum, but the text concludes that the step −ηΣ_{k∈S}g_k is a descent direction for F = Σ_k w_k L_k. Descent requires ⟨∇F(θ), Σ_{k∈S}g_k⟩ > 0, and Eq. (12) gives no control over the components of ∇F coming from inactive tasks. A singleton S={k} is always τ-compatible, and if the inactive-task gradients oppose g_k, the update can increase F. Additionally, Proposition 7 is false as stated: for g1=(1,0) and g2=(−1,0), τeff(S)=1/2 but ∥g1+g2∥²=0 < (1−1/2)(∥g1∥²+∥g2∥²)=1. The positive-part aggregate conflict ratio needs a factor of 2 in the cross-term, and the current version invalidates the 'assumption-free' refinement and its reuse in Appendix F (Eqs. 93–98).
  2. [Section 5.2/Appendix I] Eq. (13) is not derived for Algorithm 1 as run. Theorem 7 in Appendix F assumes unbiased stochastic gradients E[g_t|θ_t]=∇F(θ_t), and Appendix I.2 explicitly concedes that the deterministic cyclic schedule in Algorithm 1 generally violates this condition (Eq. 109). The proposed fixes in I.2.1 either replace the algorithm by randomized class sampling with scaling or appeal to unspecified 'standard analyses' of cyclic block updates; neither establishes Eq. (13) for the implemented method. Moreover, the (1+τ) factor in Eq. (13) does not appear in Theorem 7, and the τeff-based refinement in Appendix F bounds average active-gradient energy, not min_t E∥∇F(θ_t)∥². The claimed rate for the implemented algorithm is therefore unsupported.
  3. [Appendix C] The claim in Section 5.3 that the scheduled bound is 'never worse' than the aggregated bound is contradicted by Appendix C.4's derivation, which shows that the envelope UB_sch − UB_agg can be positive without additional structure (Eq. 54). Theorem 6's strict-improvement result is conditional on Assumption 7, which postulates negative Hessian-weighted cross-terms between separated groups—exactly the property the scheduler is meant to create rather than a property proved from τ-compatibility or graph recovery. Furthermore, the comparison is between L-smooth upper bounds, not actual objective values, so even Theorem 6 does not show that the scheduled path decreases F more than the mixed step.
  4. [Section 5.5/Appendix G] Proposition 8's bounded-staleness statement applies only while the coloring is fixed. Algorithm 1 recolors every R steps, and because R is not required to be a multiple of the current color count m_r, a task can wait across a refresh boundary for longer than Δ+1 steps. For example, with m=2 and R=3, a task in the second color class can be updated at steps 1 and 4, a gap of three steps even though Δ=1. Thus the claim that 'no task waits more than Δ steps between updates' is not guaranteed for the dynamic algorithm; the proof must either align the cycle with refresh boundaries or restate the bound for the refresh-aware schedule.
  5. [Sections 4.1–4.2] The entire grouping mechanism rests on the assumption that thresholded cosine similarity of EMA-smoothed shared-parameter gradients is a reliable proxy for the interference that actually impedes multi-task learning. The recovery theorems in Appendix B guarantee consistency with the population cosine graph under Assumptions 1–5, but they do not establish that this graph captures the interactions that slow the full objective; conflicts mediated by task-specific heads or by higher-order parameter interactions are outside the model. The paper provides no experiment validating the proxy, such as comparing graph edges against measured negative transfer or per-step objective changes, so the link between the recovered graph and the claimed performance gains remains an assumption.
minor comments (4)
  1. [Table 1] Several rows in Table 1 have formatting errors or missing separators, e.g., the SON-GOKU+AdaTask row reads '67 71 6352 0.53...' and the column header 'F&B HEALTH' conflates two datasets. The table should be reformatted and should report standard deviations or the number of seeds, since the text repeatedly claims 'consistent' improvements.
  2. [Appendix I] The sentence appealing to 'standard analyses of nonconvex smooth cyclic block updates' has no citation or theorem statement. Either provide a precise reference with the required conditions and constants, or remove the claim and state that the deterministic cyclic case is not analyzed.
  3. [General] The paper should state clearly whether Eq. (13) applies to the deterministic cyclic schedule, the randomized class-sampling variant, or both, and should define σ² as the variance of the estimator used in each case. The current presentation conflates the two variants.
  4. [General] No code or data availability statement is included, and no details of hyperparameter ranges or sensitivity (e.g., τ∗, β, R) are reported beyond the ablation. Providing these would improve reproducibility.

Circularity Check

2 steps flagged · score 5.0 of 10

Theoretical guarantees largely restate the low-conflict definition and assumed negative cross-terms; empirical benchmarks remain independent.

  1. self definitional [Section 5.1, Eq. (12); Appendix E, Proposition 6]
    "If the active set S_t at step t is τ-compatible, then the combined update is a descent direction with a quantitative lower bound: ∥Σ_{k∈S_t} g_{k,t}∥² ≥ (1−τ(|S_t|−1)) Σ_{k∈S_t}∥g_{k,t}∥² (12) ... Thus the step cannot flip to ascent whenever τ(|S_t|−1)<1. This is proved by expanding the polarization identity and controlling cross terms under the τ-compatibility condition (see Appendix E)."

    The bound is obtained by inserting the defining τ-compatibility inequality ⟨g_i,g_j⟩ ≥ −τ∥g_i∥∥g_j∥ into the polarization identity (Appendix E), so Eq. (12) is an algebraic restatement of the construction that defines a 'low-conflict' active set. Calling this 'descent' substitutes the squared norm of the active sum for the required inner product ⟨∇F, Σ_{k∈S} g_k⟩; the full objective F also contains inactive-task gradients, which are unconstrained by Eq. (12). The descent guarantee is therefore not an independent consequence for the full MTL objective—it is the definition of τ-compatibility repackaged as a descent statement.

  2. self definitional [Section 5.3; Appendix C.5, Assumption 7 and Theorem 6]
    "Using a telescoping L-smooth bound and evaluating both trajectories at a common linearization—i.e., expanding F at the refresh start θ_tr and applying the same first-order model with the same step size—the scheduled bound is never worse and is strictly better when cross-group interaction terms are sufficiently negative (so mixed updates would cancel progress). ... Assumption 7: There exist nonnegative margins {Γ_pq}_{p<q} such that I_pq(x;η) := ∫_0^1 ⟨H(x−tηG^0_p)G^0_p, G^0_q⟩ dt ≤ −Γ_pq∥G^0_p∥∥G^0_q∥ for all p<q."

    Theorem 6's strict improvement (UB_sch < UB_agg) is driven by the assumed negative Hessian-weighted cross-term Γ_pq, which is exactly the claim that separated groups have the negative interaction the scheduler is designed to detect and separate. The theorem does not establish that the scheduler produces such cross-terms; it assumes the phenomenon as its premise, so the 'scheduled beats mixed' prediction is equivalent to the input assumption rather than a derived consequence.

full rationale

The paper's empirical evaluation against six external datasets is independent and not circular, and there is no load-bearing self-citation chain: the graph-coloring and concentration facts cited are standard external results. However, the central theoretical predictions are partially circular by construction. Section 5.1 proves descent by lower-bounding ∥Σ_{k∈S} g_k∥² under the very pairwise-cosine definition used to build the groups; for the full MTL objective F = Σ w_k L_k, descent requires an inner product with the full gradient, which includes inactive tasks, so the claimed full-objective descent is not derived—it is the low-conflict definition relabeled. The 'scheduled beats mixed' theorem in Section 5.3 is conditional on Assumption 7, whose content is that separated groups have negative Hessian-weighted cross-terms, precisely the interference the scheduler is designed to enforce; the strict improvement term is the assumed negative cross-term itself. I do not count Eq. (13) as circularity: the nonconvex-rate claim is unsupported for the actual deterministic cyclic algorithm because Appendix I's unbiased-SGD theorem (Theorem 8) does not apply to the biased update in Eq. (109), and the repair in Appendix I.2 switches to randomized scaled class sampling rather than Algorithm 1. That is a support gap, not a definitional reduction. Overall, the theoretical 'predictions' that grouping preserves descent and beats mixed updates reduce to definitions or to the exact assumed negative interference, so partial circularity is present; score 5 reflects that the empirical content remains non-circular.

Assumptions & free parameters 5 free parameters · 9 assumptions · 0 invented entities

The main theory imports standard regularity conditions but adds several ad hoc assumptions that are either not verified on the benchmarks or, in the case of Assumption 7, encode the conclusion being proved. The empirical method is not fitted to the test sets, but the theoretical guarantees are weaker than the prose suggests.

free parameters (5)
  • Conflict threshold tau target = not reported (annealed from 1 to tau*)
    Controls graph density, group sizes, and update frequency; user-chosen in Section 4.4.
  • EMA smoothing coefficient beta = not reported
    Controls effective sample size neff and noise in cosine estimates; introduced in Section 4.1.
  • Refresh period R = main results unspecified; wall-clock study uses R in {4, 32, 256}
    Determines how often the conflict graph is rebuilt and the amortized scheduler cost K^2 d / R.
  • Warm-up steps T_warm = not reported
    Sets tau = 1 for the first T_warm steps to mitigate noisy early gradients; part of Section 4.4.2.
  • Minimum coverage f_min = not reported
    Duplicates singleton tasks into compatible steps to guarantee update frequency; introduced in Section 4.4.1.
assumptions (9)
  • standard math The multi-task objective F is L-smooth.
    Used throughout Appendix C and F to apply the standard descent inequality; typical for neural networks but not verified for the specific models.
  • domain assumption Stochastic gradients are unbiased for the full objective.
    Equation (13) and Theorem 7 assume E[g_t | theta_t] = grad F(theta_t). Algorithm 1's deterministic cyclic schedule violates this, and Appendix I patches it with randomized class sampling with scaling, which is not the algorithm evaluated.
  • ad hoc to paper Separation margin gamma exists around the conflict threshold.
    Assumption 1 in Appendix B; exact recovery of the population conflict graph requires population cosines to be at least gamma away from -tau, which is not verified empirically.
  • domain assumption Population task gradients drift slowly within a refresh window.
    Assumption 3 in Appendix B; needed for EMA concentration to keep cosine estimates close to start-of-window values.
  • domain assumption Every included task gradient has norm at least m0.
    Assumption 4 in Appendix B; needed for cosine perturbation bounds. The implementation excludes low-norm tasks using a threshold nu.
  • domain assumption Gradient noise is sub-Gaussian and satisfies mixing or martingale-difference conditions.
    Assumption 2 in Appendix B; needed for the exponential-weight concentration bounds that drive the recovery guarantees.
  • domain assumption Threshold tau is fixed or selected on data independent of the probe window.
    Assumption 5 in Appendix B; avoids selection effects in the recovery argument but is not discussed in the experimental setup.
  • ad hoc to paper Per-group Lipschitzness of the group gradient maps.
    Assumption 6 in Appendix C; used to compare scheduled and aggregated upper bounds. The authors argue it follows from task-level smoothness, but it is not verified for the trained networks.
  • ad hoc to paper Negative Hessian-weighted cross-terms between separated groups.
    Assumption 7 in Appendix C.5; this is the load-bearing assumption for Theorem 6 because the proof of scheduled-beats-mixed only goes through when separated groups have negative Hessian-weighted interactions, which is exactly the property the scheduler is designed to create.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Graph Coloring for Multi-Task Learning." pith.science (2026). https://pith.science/paper/HOXTUU35

@misc{pith2026250916959,
  author       = {Pith},
  title        = {Pith review of: Graph Coloring for Multi-Task Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HOXTUU35}},
  note         = {Machine review of arXiv:2509.16959}
}
read the original abstract

When different objectives conflict with each other in multi-task learning, gradients begin to interfere and slow convergence, thereby potentially reducing the final model's performance. To address this, we introduce SON-GOKU, a scheduler that computes gradient interference, constructs an interference graph, and then applies greedy graph-coloring to partition tasks into groups that align well with each other. At each training step, only one group (color class) of tasks are activated, and the grouping partition is constantly recomputed as task relationships evolve throughout training. By ensuring that each mini-batch contains only tasks that pull the model in the same direction, our method improves the effectiveness of any underlying multi-task learning optimizer without additional tuning. Since tasks within these groups will update in compatible directions, multi-task learning will improve model performance rather than impede it. Empirical results on six different datasets show that this interference-aware graph-coloring approach consistently outperforms baselines and state-of-the-art multi-task optimizers. We provide extensive theory showing why grouping and sequential updates improve multi-task learning, with guarantees on descent, convergence, and the ability to accurately identify what tasks conflict or align.

Figures

Figures reproduced from arXiv: 2509.16959 by the authors.

Figure 1
Figure 1. Interference-aware scheduling pipeline: (a) For each task [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 11 canonical work pages

  1. [1]

    Uniform.This baseline assigns equal weights to all tasks throughout training, representing the simplest approach where all task losses are weighted equally

    work page

  2. [2]

    This maintains consistent training dynamics across tasks

    Gradnorm (Chen et al., 2018).Balances task learning rates by normalizing gradient magnitudes relative to target loss ratios. This maintains consistent training dynamics across tasks

    work page 2018

  3. [3]

    odd” element among “even

    URLhttps://arxiv.org/abs/1806.08730. Florence Merlevède, Magda Peligrad, and Emmanuel Rio. Bernstein inequality and moderate deviations under strong mixing conditions.High Dimensional Probability VI, pp. 273–292, 2011. Kaisa Miettinen.Nonlinear Multiobjective Optimization. Kluwer Academic Publishers, Boston, MA, 1999. ISBN 978-0792382781. Tomas Mikolov, K...

    arXiv 2011

  4. [4]

    It uses an online, per-step rule (no pairwise gradient ops), adding negligible overhead while remaining robust to loss-scale differences

    FAMO (Liu et al., 2023).Fast Adaptive Multitask Optimization dynamically adjusts task weights to equalize each task’s rate of loss improvement. It uses an online, per-step rule (no pairwise gradient ops), adding negligible overhead while remaining robust to loss-scale differences

    work page 2023

  5. [5]

    The parameterα controls the trade-off between average performance and fairness

    Fair Resource Allocation in MTL (FairGrad) (Ban & Ji, 2024).Views the shared update as a limited resource and chooses it to maximize anα-fair utility of per-task improvements. The parameterα controls the trade-off between average performance and fairness

    work page 2024

  6. [6]

    Weights are obtained by solving a small inner problem (e.g., via CCP) using the gradient Gram matrix

    Nash-MTL (Navon et al., 2022).Frames multitask training as a bargaining game and computes a scale-invariant weighted combination of task gradients given by the Nash bargaining solution. Weights are obtained by solving a small inner problem (e.g., via CCP) using the gradient Gram matrix. Updates are balanced across tasks. L Expanded W all-Clock Time Study ...

    work page arXiv 2020

  7. [8]

    MGDA (Sener & Koltun, 2018).Formulates multi-task learning as a multi-objective optimization problem, finding Pareto-optimal solutions (Lockwood, 2008; Pareto,

    work page 2018

  8. [10]

    PCGrad (Yu et al., 2020).Projects conflicting gradients onto orthogonal subspaces when negative cosine similarity is detected, eliminating destructive interference between task gradients

    work page 2020

Show all 15 references
  1. [11]

    This proves more nuanced modifications to gradients than binary projection

    CAGrad (Liu et al., 2021).Extends PCGrad by adaptively adjusting gradient magnitudes based on conflict severity. This proves more nuanced modifications to gradients than binary projection

    2021

  2. [12]

    Adatask (Yang et al., 2023).Dynamically reweighs task losses using relative loss changes, adapting to varying task learning rates during training

    2023

  3. [2004]

    Transfer learning: even/odd MNIST

    ISBN 978-1-4419-8853-9. Feng Niu, Benjamin Recht, Christopher Ré, and Stephen J. Wright. Hogwild!: A lock-free approach to parallelizing stochastic gradient descent. InAdvances in Neural Information Processing Systems, volume 24, pp. 693–701, 2011. 13 Vilfredo Pareto.Manual of...

    2011 arXiv

  4. [2014]

    36 K.2 State-of-the-art models

    through gradient descent in the convex hull of gradients (Fliege & Svaiter, 2000; Miettinen, 1999). 36 K.2 State-of-the-art models

    2000

  5. [2015]

    J.2 CIF AR-10 The CIFAR-10 (Krizhevsky et al., 2009) dataset contains 60,00032× 32color images across 10 generic classes

    backbone trained from scratch processes the RGB input, with task-specific decoder heads for segmentation (with 32× upsampling), depth regression, surface normal regression, and color temperature estimation. J.2 CIF AR-10 The CIFAR-10 (Krizhevsky et al., 2009) dataset contains ...

    2009

  6. [2016]

    Valerio Guarrasi, Fatih Aksu, Camillo Maria Caruso, Francesco Di Feola, Aurora Rofena, Filippo Ruffini, and Paolo Soda

    doi: 10.1137/15M1009718. Valerio Guarrasi, Fatih Aksu, Camillo Maria Caruso, Francesco Di Feola, Aurora Rofena, Filippo Ruffini, and Paolo Soda. A systematic review of intermediate fusion in multimodal deep learning for biomedical applications.Image and Vision Computing, pp. 1...

    arXiv

  7. [2018]

    Sanjoy Dasgupta and Anupam Gupta

    URLhttp://arxiv.org/abs/1801.02143. Sanjoy Dasgupta and Anupam Gupta. An elementary proof of a theorem of johnson and lindenstrauss. InRandomization and Approximation Techniques in Computer Science, RANDOM 2003, pp. 53–62. Springer, 2003. doi: 10.1007/978-3-540-45198-3_4. Vict...

    2003 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.