pith. sign in

arxiv: 1907.08600 · v1 · pith:6NN4UUECnew · submitted 2019-07-19 · 💻 cs.LG · stat.ML

Learning sparsity in reservoir computing through a novel bio-inspired algorithm

Luca Manneschi , Andrew C. Lin , Eleni Vasilaki This is my paper

Pith reviewed 2026-05-24 19:06 UTC · model grok-4.3

classification 💻 cs.LG stat.ML
keywords reservoir computingsparsitybio-inspiredfiring thresholdsgradient descentMarkov chain Monte Carlomushroom bodyclassification
0
0 comments X

The pith

A bio-inspired algorithm learns optimal sparsity in reservoir computing by tuning neuron firing thresholds.

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

This paper presents a method for optimizing sparsity in reservoir computing models, drawing from the sparse activity in the fruit fly's mushroom body used for odor classification. The approach uses gradient descent to learn neuron-specific firing thresholds and a Markov chain Monte Carlo method for a global threshold, but applies sparsity only to the readout layer. This allows the model to achieve better classification performance, memorization, and faster convergence compared to standard gradient descent on readout weights alone. A sympathetic reader would care because it offers a simple way to incorporate biological sparsity principles into machine learning without disrupting the reservoir's internal dynamics.

Core claim

The paper claims that by taking inspiration from the inhibitory feedback and high firing thresholds in the fruit fly mushroom body, a hybrid algorithm of gradient descent and Markov chain Monte Carlo can optimize sparsity in the readout layer of a reservoir, outperforming standard methods on two example tasks and improving classification, memorization ability, and convergence time.

What carries the argument

The hybrid learning rule for firing thresholds: neuron-specific thresholds updated via gradient descent combined with a global threshold via Markov chain Monte Carlo, restricted to the readout layer to preserve reservoir timescales.

If this is right

  • The learnt sparse representation leads to better classification performance on tasks.
  • It improves the memorization ability of the model.
  • Convergence time is reduced compared to standard gradient descent.
  • The algorithm can be derived as a one-layer update rule due to the readout-only application.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This approach might extend to other recurrent network models where sparsity could aid efficiency.
  • Similar mechanisms could be explored in hardware implementations to reduce energy use.
  • Further tasks beyond the two examples could test the generalizability of the performance gains.

Load-bearing premise

The sparsity is only applied on the readout layer so as not to change the timescales of the reservoir and to allow the derivation of a one-layer update rule for the firing thresholds.

What would settle it

Running the proposed model and standard gradient descent on the two example tasks and finding no improvement in performance, memorization, or convergence time would falsify the outperformance claim.

Figures

Figures reproduced from arXiv: 1907.08600 by Andrew C. Lin, Eleni Vasilaki, Luca Manneschi.

Figure 1
Figure 1. Figure 1: Scheme of the network and of the tasks considered. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Task 1, performance. Left. Performance of the model during training. The Composed model has the highest speed of convergence, while the model without thresholds and sparse activity has lowest accuracy. Right Fraction of correct classifications after 60000 episodes. The difference between the algorithms increases with the difficulty of the task [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Task 1, Mean and variance of the optimized θ distribution. Left The average of the distribution shows an upward trend and the sparseness in the network rises with the number of input stimuli. Right The need to differentiate the values of the thresholds is reflected in the σ of the optimized distribution and it increases as the task becomes more demanding. The surprising results obtained by optimizing a glo… view at source ↗
Figure 5
Figure 5. Figure 5: fig.5. The [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Task 1, performance. Left. Performance of the model during training. The performance of the composed model are the highest. The most remarkable difference in comparison to the results of Task 1 is the low accuracy reported by the Metropolis, which is the model that optimizes a global threshold. Since the task considered is dynamic, this result is expected (see text) [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Task 2, Mean and variance of the optimized θ distribution. Even if there is no evident trend in the average or the variance of the distribution, the results showed in fig.3 are robust with respect to the variations of the values showed. GDθ reports the highest mean, which is suboptimal if the performance of fig.2 are considered. Thus, the presence of the global threshold in the Composed model helps the mod… view at source ↗
Figure 6
Figure 6. Figure 6: Change in the level of specificity after training. Left. Distribution of Spi before learning. Right. Distribution of Spi after learning [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Performance of the proposed algorithm in a Reinforcement Learning framework. gle global parameter for the whole reservoir. It would be possible to adopt this random perturba￾tion to each single neuron separately, but the algo￾rithm would have to learn a high number of pa￾rameters through stochastic search, which would dramatically increase the convergence time, while the learning of a global threshold allo… view at source ↗
read the original abstract

The mushroom body is the key network for the representation of learned olfactory stimuli in Drosophila and insects. The sparse activity of Kenyon cells, the principal neurons in the mushroom body, plays a key role in the learned classification of different odours. In the specific case of the fruit fly, the sparseness of the network is enforced by an inhibitory feedback neuron called APL, and by an intrinsic high firing threshold of the Kenyon cells. In this work we took inspiration from the fruit fly brain to formulate a novel machine learning algorithm that is able to optimize the sparsity level of a reservoir by changing the firing thresholds of the nodes. The sparsity is only applied on the readout layer so as not to change the timescales of the reservoir and to allow the derivation of a one-layer update rule for the firing thresholds. The proposed algorithm is a combination of learning a neuron-specific sparsity threshold via gradient descent and a global sparsity threshold via a Markov chain Monte Carlo method. The proposed model outperforms the standard gradient descent, which is limited to the readout weights of the reservoir, on two example tasks. It demonstrates how the learnt sparse representation can lead to better classification performance, memorization ability and convergence time.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a bio-inspired algorithm for reservoir computing that learns sparsity by optimizing neuron-specific firing thresholds via gradient descent and a global sparsity threshold via MCMC. Sparsity is restricted to the readout layer to preserve reservoir timescales and enable a one-layer update rule. The central claim is that this approach outperforms standard gradient descent (limited to readout weights) on two example tasks, yielding better classification performance, memorization ability, and convergence time.

Significance. If the outperformance claims hold under rigorous validation, the work provides a concrete translation of Drosophila mushroom-body sparsity mechanisms (APL inhibition and high Kenyon-cell thresholds) into an RC training procedure. The hybrid GD+MCMC scheme for threshold learning is a distinctive contribution that could improve efficiency in echo-state networks without altering internal dynamics.

major comments (2)
  1. [Abstract] Abstract: the claim that the proposed model 'outperforms the standard gradient descent... on two example tasks' is load-bearing for the central contribution, yet the abstract (and by extension the results presentation) supplies no quantitative metrics, error bars, baseline details, dataset descriptions, or statistical tests, leaving the strength of support for the claim unclear.
  2. [Methods / Algorithm Description] The modeling choice to restrict sparsity to the readout layer is justified in the abstract as enabling a one-layer update rule, but the manuscript must explicitly derive or demonstrate that the combined GD+MCMC updates do not inadvertently alter reservoir timescales or introduce hidden dependencies on the free parameters (neuron-specific thresholds and global sparsity threshold).
minor comments (2)
  1. [Algorithm] Clarify the precise form of the one-layer update rule for firing thresholds and state whether it remains parameter-free after the MCMC step.
  2. [Experiments] Provide full experimental protocols, including reservoir size, spectral radius, task definitions, and exact baseline implementations, to allow reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and indicate the revisions we will make.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that the proposed model 'outperforms the standard gradient descent... on two example tasks' is load-bearing for the central contribution, yet the abstract (and by extension the results presentation) supplies no quantitative metrics, error bars, baseline details, dataset descriptions, or statistical tests, leaving the strength of support for the claim unclear.

    Authors: We agree that the abstract would be strengthened by quantitative support. In the revised manuscript we will expand the abstract to report specific performance metrics (e.g., accuracy or error reductions), error bars from repeated trials, brief dataset and baseline descriptions, and reference to the statistical tests used. Corresponding details will also be added to the results section. revision: yes

  2. Referee: [Methods / Algorithm Description] The modeling choice to restrict sparsity to the readout layer is justified in the abstract as enabling a one-layer update rule, but the manuscript must explicitly derive or demonstrate that the combined GD+MCMC updates do not inadvertently alter reservoir timescales or introduce hidden dependencies on the free parameters (neuron-specific thresholds and global sparsity threshold).

    Authors: We will add an explicit derivation in a new methods subsection. Because sparsity is applied exclusively after the reservoir state is generated, the reservoir recurrence and timescales remain untouched; the GD step updates only the per-neuron readout thresholds and the MCMC step updates only the scalar global threshold. We will show the resulting one-layer update equations and confirm that no gradients or proposals propagate back into the reservoir weights or dynamics, thereby excluding hidden dependencies. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper's central algorithm combines gradient descent on neuron-specific firing thresholds with MCMC on a global sparsity threshold, restricted to the readout layer by explicit modeling choice to preserve reservoir timescales and permit a one-layer update. No equation or claim reduces by construction to a fitted parameter renamed as prediction, nor to a self-citation chain; the outperformance versus readout-only GD is presented as an empirical result on example tasks rather than a definitional identity. The derivation is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on the domain assumption that Drosophila-style sparsity mechanisms can be directly ported to reservoir readout layers without disrupting dynamics, plus the untested premise that the combined GD+MCMC procedure yields superior performance on unseen tasks.

free parameters (2)
  • neuron-specific firing thresholds
    These are learned parameters whose values are fitted during training; their number equals the number of readout nodes.
  • global sparsity threshold
    Learned via MCMC; acts as a hyperparameter controlling overall sparsity level.
axioms (2)
  • domain assumption Sparse activity of Kenyon cells enforced by APL inhibition and high firing threshold is key to learned odor classification in Drosophila
    Invoked in the abstract as the biological basis for the algorithm design.
  • domain assumption Applying sparsity only to the readout layer preserves reservoir timescales
    Stated explicitly as the reason a one-layer update rule can be derived.

pith-pipeline@v0.9.0 · 5739 in / 1527 out tokens · 37686 ms · 2026-05-24T19:06:05.217447+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages · 1 internal anchor

  1. [1]

    Sparseness of the neuronal representation of stimuli in the primate temporal visual cortex

    Edmund T Rolls and Martin J Tovee. Sparseness of the neuronal representation of stimuli in the primate temporal visual cortex. Journal of neurophysiology , 73(2):713–726, 1995

    work page 1995

  2. [2]

    Sensory processing in the drosophila anten- nal lobe increases reliability and separability of en- semble odor representations

    Vikas Bhandawat, Shawn R Olsen, Nathan W Gouwens, Michelle L Schlief, and Rachel I Wil- son. Sensory processing in the drosophila anten- nal lobe increases reliability and separability of en- semble odor representations. Nature neuroscience, 10(11):1474, 2007

    work page 2007

  3. [3]

    Sparse, decorrelated odor coding in the mushroom body enhances learned odor discrimination

    Andrew C Lin, Alexei M Bygrave, Alix De Calignon, Tzumin Lee, and Gero Miesenb¨ ock. Sparse, decorrelated odor coding in the mushroom body enhances learned odor discrimination. Nature neuroscience, 17(4):559, 2014

    work page 2014

  4. [4]

    Learning with structured sparsity

    Junzhou Huang, Tong Zhang, and Dimitris Metaxas. Learning with structured sparsity. Jour- nal of Machine Learning Research , 12(Nov):3371– 3412, 2011

    work page 2011

  5. [5]

    Statistical learning with sparsity: the lasso and generalizations

    Trevor Hastie, Robert Tibshirani, and Martin Wainwright. Statistical learning with sparsity: the lasso and generalizations. Chapman and Hall/CRC, 2015

    work page 2015

  6. [6]

    Enhancing sparsity by reweighted l 1 minimization

    Emmanuel J Candes, Michael B Wakin, and Stephen P Boyd. Enhancing sparsity by reweighted l 1 minimization. Journal of Fourier analysis and applications, 14(5-6):877–905, 2008

    work page 2008

  7. [7]

    Learning structured sparsity in deep neural networks

    Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Learning structured sparsity in deep neural networks. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems 29, pages 2074–2082. Curran Associates, Inc., 2016

    work page 2074

  8. [8]

    Dropout: a simple way to prevent neural networks from overfitting

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhut- dinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929–1958, 2014

    work page 1929

  9. [9]

    Model sparsity and brain pattern inter- pretation of classification models in neuroimaging

    Peter M Rasmussen, Lars K Hansen, Kristoffer H Madsen, Nathan W Churchill, and Stephen C Strother. Model sparsity and brain pattern inter- pretation of classification models in neuroimaging. Pattern Recognition, 45(6):2085–2100, 2012

    work page 2085

  10. [10]

    Optimization and applications of echo state networks with leaky- integrator neurons

    Herbert Jaeger, Mantas Lukoˇ seviˇ cius, Dan Popovici, and Udo Siewert. Optimization and applications of echo state networks with leaky- integrator neurons. Neural networks, 20(3):335–352, 2007

    work page 2007

  11. [11]

    A connectome of a learning and memory center in the adult drosophila brain

    Shin-ya Takemura, Yoshinori Aso, Toshihide Hige, Allan Wong, Zhiyuan Lu, C Shan Xu, Patricia K Rivlin, Harald Hess, Ting Zhao, Toufiq Parag, et al. A connectome of a learning and memory center in the adult drosophila brain. Elife, 6:e26975, 2017

    work page 2017

  12. [12]

    Gap junction networks in mushroom bodies participate in visual learning and memory in drosophila

    Qingqing Liu, Xing Yang, Jingsong Tian, Zhong- bao Gao, Meng Wang, Yan Li, and Aike Guo. Gap junction networks in mushroom bodies participate in visual learning and memory in drosophila. Elife, 5:e13238, 2016

    work page 2016

  13. [13]

    Random convergence of olfactory inputs in the drosophila mushroom body

    Sophie JC Caron, Vanessa Ruta, LF Abbott, and Richard Axel. Random convergence of olfactory inputs in the drosophila mushroom body. Nature, 497(7447):113, 2013

    work page 2013

  14. [14]

    The echo state approach to analysing and training recurrent neural networks- with an erratum note

    Herbert Jaeger. The echo state approach to analysing and training recurrent neural networks- with an erratum note. Bonn, Germany: German National Research Center for Information Technol- ogy GMD Technical Report , 148(34):13, 2001

    work page 2001

  15. [15]

    Tutorial on training recur- rent neural networks, covering BPPT, RTRL, EKF and the” echo state network” approach , vol- ume 5

    Herbert Jaeger. Tutorial on training recur- rent neural networks, covering BPPT, RTRL, EKF and the” echo state network” approach , vol- ume 5. GMD-Forschungszentrum Informationstech- nik Bonn, 2002

    work page 2002

  16. [16]

    Coding of odors by a receptor repertoire

    Elissa A Hallem and John R Carlson. Coding of odors by a receptor repertoire. Cell, 125(1):143– 160, 2006

    work page 2006

  17. [17]

    Divisive normalization in olfactory popula- tion codes

    Shawn R Olsen, Vikas Bhandawat, and Rachel I Wilson. Divisive normalization in olfactory popula- tion codes. Neuron, 66(2):287–299, 2010

    work page 2010

  18. [18]

    Gen- erating sparse and selective third-order responses in the olfactory system of the fly

    Sean X Luo, Richard Axel, and LF Abbott. Gen- erating sparse and selective third-order responses in the olfactory system of the fly. Proceedings of the National Academy of Sciences , 107(23):10713– 10718, 2010. 12

    work page 2010

  19. [19]

    Odor discrimination in drosophila: from neural population codes to behav- ior

    Moshe Parnas, Andrew C Lin, Wolf Huetteroth, and Gero Miesenb¨ ock. Odor discrimination in drosophila: from neural population codes to behav- ior. Neuron, 79(5):932–944, 2013

    work page 2013

  20. [20]

    Disorder and the neural representation of complex odors: smelling in the real world

    Kamesh Krishnamurthy, Ann M Hermundstad, Thierry Mora, Aleksandra M Walczak, and Vijay Balasubramanian. Disorder and the neural repre- sentation of complex odors: smelling in the real world. arXiv preprint arXiv:1707.01962 , 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  21. [21]

    Monte carlo as- sessment of parameter uncertainty in conceptual catchment models: the metropolis algorithm

    George Kuczera and Eric Parent. Monte carlo as- sessment of parameter uncertainty in conceptual catchment models: the metropolis algorithm. Jour- nal of Hydrology , 211(1-4):69–85, 1998

    work page 1998

  22. [22]

    Reservoir computing approaches to recurrent neural network training

    Mantas Lukoˇ seviˇ cius and Herbert Jaeger. Reservoir computing approaches to recurrent neural network training. Computer Science Review , 3(3):127–149, 2009

    work page 2009

  23. [23]

    Learning curves for stochastic gradient descent in linear feedforward networks

    Justin Werfel, Xiaohui Xie, and H Sebastian Seung. Learning curves for stochastic gradient descent in linear feedforward networks. In Advances in neural information processing systems , pages 1197–1204, 2004

    work page 2004

  24. [24]

    Reinforcement signalling in drosophila; dopamine does it all after all

    Scott Waddell. Reinforcement signalling in drosophila; dopamine does it all after all. Current opinion in neurobiology , 23(3):324–329, 2013

    work page 2013

  25. [25]

    Sparse representa- tion for signal classification

    Ke Huang and Selin Aviyente. Sparse representa- tion for signal classification. In Advances in neu- ral information processing systems , pages 609–616, 2007. 13

    work page 2007