{"id":"93798bb6-0a8c-411a-a388-928210913241","arxiv_id":"2501.06226","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"asanAI is an open-source, browser-based toolkit that lets non-coders design and train neural networks offline, with visualizations and code export.","lead":"This paper introduces asanAI, a no-code, offline-first machine learning toolkit that runs in a web browser and lets non-programmers build, train, and test neural networks. A smart generalist should look at it as an example of accessibility-driven ML education software that could lower classroom and prototyping entry barriers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central accessibility claim rests on an untested interactivity-learning assumption; no user study or usability data is reported.","rationale":"The reader's weakest assumption is exactly the one I find most load-bearing: the Interactivity section asserts, without support, that immediate visual feedback produces higher engagement and better retention, and the paper's educational-value conclusion depends on this assertion. I agree with the conditional verdict because the paper is a clear, plausible feature description, but its central empirical claim is unvalidated. I also note the footnote in the Introduction that \"full features\" offline excludes loading example data, which slightly weakens the offline-first claim but is not the primary concern. No code repository, demo evaluation, or formal verification is included, so the toolkit's described functionality is not independently confirmed; however, I am not treating that as a correctness flaw, only as a reason for the conditional status. The proposed user study directly tests the causal mechanism the paper asserts and would either provide the missing support or force a more modest scoping of the contribution.","tokens_in":7478,"tokens_out":2884,"duration_ms":32740,"concrete_test":"Pre-register and run a randomized between-subjects user study with at least 30 participants who have no ML or programming background. One group completes an image-classification task with asanAI; the control group uses a minimal equivalent baseline, such as Teachable Machine or a guided Python notebook. Measure task completion rate, time-on-task, error rate, System Usability Scale, and pre/post ML-concept quiz scores. Set a priori thresholds: mean SUS >= 68 and post-test learning gain significantly higher than control (p < 0.05, Cohen's d >= 0.5). If these thresholds are not met, the central claim about lowering educational barriers must be downgraded to a description of interface features rather than demonstrated impact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that asanAI \"has lowered the barriers to entry\" for people \"regardless of their technical expertise or resources\" and that it is an effective educational tool. The only stated mechanism connecting the GUI to learning outcomes is in Design Concepts, Interactivity: \"Interactive software with immediate response ... leads to higher engagement levels and better information retention.\" This is asserted without a citation, a prior literature link, or an experiment. The remainder of the paper is a feature description: no user study, no usability metrics, no baseline comparison to Teachable Machine, ml5.js, Lobe, or a simple Python notebook, and no measurement of learning gains. The abstract's statement that the toolkit is \"successfully utilized\" by researchers, trainers, and teachers is also unsupported by any reported data. This concern is load-bearing because if the assumed engagement/retention benefit does not materialize, the central democratizing-ML contribution collapses to \"a GUI exists,\" which existing tools already provide. I am not claiming the toolkit is non-functional; I am claiming the proof burden for the paper's empirical educational claim is unmet. A software-description paper could be acceptable if scoped as such, but the conclusion makes a causal claim about lowered barriers that needs evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript describes asanAI, an open-source, no-code, browser-based machine learning toolkit that lets users design, train, debug, and test neural networks directly in a web browser. The paper emphasizes offline-first operation, local computation for privacy, WebGL acceleration, touch-friendly interfaces, automatic shape detection, a range of visualizations (FCNN, LeNet, feature maps, GradCAM, and Math-mode), data importers for CSV/images/tensors, and export of models and generated Python code. The stated goal is to lower entry barriers to machine learning and support teaching of ML in school and hobbyist contexts. The conclusion claims that the tool has been successful in lowering these barriers and is already used by researchers, trainers, and teachers, but the manuscript contains no user study, no usability metrics, no benchmarks, and no quantitative evidence for these claims.","tokens_in":7677,"tokens_out":4843,"duration_ms":49712,"significance":"If the effectiveness claims were demonstrated, asanAI would occupy a genuinely useful niche among educational ML toolkits by combining offline operation, no-code touch-friendly interaction, privacy through local computation, open-source licensing, and export to standard formats. The paper is strongest in documenting the feature set and design rationale, and the open-source artifact is a concrete contribution. The manuscript's significance is currently limited, however, because the central claims about learning outcomes, engagement, and barrier-lowering are asserted rather than evidenced. The paper would be acceptable as a software description if re-scoped, or as a strong system paper if accompanied by an evaluation; in its present form, the gap between the claims and the evidence is substantial.","major_comments":[{"comment":"The paragraph states that interactive software with immediate response 'leads to higher engagement levels and better information retention' and uses this as a core justification for the toolkit's educational value. This is a load-bearing, causal claim that is made without a citation, prior literature, or any experimental data. The paper does not report a user study measuring learning gains, engagement, or retention, nor any usability metrics. Please either supply relevant prior-work citations, add an empirical evaluation, or explicitly re-scope the paper as a system description that does not make claims about learning outcomes.","section":"Design Concepts, Interactivity"},{"comment":"The abstract asserts that the toolkit is 'successfully utilized by researchers of ScaDS.AI,' by trainers, and by teachers, and the conclusion asserts that the work 'has lowered the barriers to entry' for users regardless of technical expertise or resources. No data support these statements: there is no usage log, survey, interview, case study, benchmark, or comparison with existing tools such as Teachable Machine, ml5.js, or Lobe. If these claims are to remain, they require supporting evidence; otherwise, the manuscript should be reformulated around capabilities and intended uses rather than demonstrated effectiveness.","section":"Abstract and Conclusion and Future Works"},{"comment":"There is an internal contradiction regarding offline operation. The abstract says full offline use is possible, and the Introduction states that once the page is loaded, 'the internet connection is no longer required, and its full features can be safely used offline,' yet footnote 3 says 'With the exception of loading example data.' This matters because offline-first is a central design pillar. Please clarify which features (including example data, tutorials, and documentation) require connectivity and adjust the abstract and Introduction accordingly.","section":"Introduction, footnote 3 and Abstract"},{"comment":"The comparison with Teachable Machine is inaccurate: the paper states that Teachable Machine 'requires an active and reliable Internet connection, as all computations are performed on the server side.' The cited work (Carney et al. 2020) describes Teachable Machine as a browser-based tool, and the widely available Teachable Machine 2 performs inference and training in the browser via TensorFlow.js, with data staying on-device. This mischaracterization affects the paper's differentiation claims and should be corrected.","section":"Related work"}],"minor_comments":[{"comment":"The sentence 'For CSV data, the input shape is set to [nr of X columns], where nr of X columns is For CSV data...' is malformed and incomplete; it should be rewritten to state the input and output shapes clearly.","section":"Implementation Considerations"},{"comment":"There are typos and inconsistent terms, e.g., 'Ternsorflow' instead of 'TensorFlow,' and 'asanai.js' vs. 'asanAI' capitalization; a careful proofreading pass is needed.","section":"Throughout"},{"comment":"The statement that exported data is provided under 'CC-BY 4' is ambiguous: it is unclear whether this license applies to the user's exported model and data or only to the tool's templates and defaults. Please clarify.","section":"Interoperability"},{"comment":"The paper says asanAI is released as open-source software under an MIT-like license but does not provide a link to the source repository or a version identifier. Please add a repository URL and, ideally, a version/commit reference to support reproducibility.","section":"Introduction"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is essentially a feature description with strong claims about usability and educational impact but no evaluation. I would strongly recommend either adding a rigorous user study or explicitly re-scoping the paper as a software/demo contribution. The Related Work inaccuracy about Teachable Machine should be corrected, as it currently overstates the novelty and could mislead readers. If the target venue has a software-track with different evidence standards, the paper might fit there; as submitted for a general ML journal, the evidence gap is too large for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Brief: This is a decent software description with an unmet evaluation burden. The toolkit itself looks genuinely useful. The paper's claims about learning outcomes, 'lowered barriers,' and 'successfully utilized' are asserted, not demonstrated.\n\nWhat's new: the combination of offline-first, no-code, touch-friendly browser interface with GradCAM, feature maps, and a math mode that shows the actual equations. That's a nice set of features, and the paper does a good job of walking through the GUI, the operational modes, and the implementation details. It cites the relevant prior work fairly and is honest about future directions.\n\nThe soft spots are real. There is no user study, no usability metrics, no benchmark, and no baseline comparison to Teachable Machine, Lobe, ml5.js, or a simple Python notebook. The abstract says asanAI is 'successfully utilized' by researchers, trainers, and teachers, but no data backs that up. The design section claims interactive interfaces lead to 'higher engagement levels and better information retention' with no citation. The offline-first claim is also slightly contradicted by footnote 3, which admits example data cannot be loaded without a connection. And while the paper says the code is open source, it does not provide a repository link, so a reviewer cannot easily verify the implementation.\n\nNone of this means the tool does not work. But the paper's conclusion overreaches: 'has lowered barriers' is a causal claim the paper does not support. If the authors resubmit with a real usability evaluation, even a small one, the contribution would be credible. Alternatively, they could honestly scope the paper as a tool announcement and stop making empirical claims.\n\nFor peer review: yes, this deserves referee time, but only with the expectation of major revisions. A serious referee would ask for the missing evaluation or a significant toning down of the claims. I would send it out rather than desk reject, because the tool has educational potential and the field benefits from this kind of work being documented.","headline":"A useful software description whose central educational claims are asserted, not shown; send it to review with expectations of major revision.","tokens_in":8173,"tokens_out":2549,"would_cite":false,"duration_ms":24411,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that asanAI—a no-code, offline-first browser toolkit—can let anyone without programming skills design, train, and test neural networks on the device they already own.","keywords":["asanAI","no-code machine learning","browser-based training","offline-first computing","neural network visualization","machine learning education","model export"],"falsifier":"A controlled experiment with two matched groups of learners—one building and training networks in asanAI, one receiving the same concepts through a static slide-based lesson—with a delayed test of concept understanding would settle the educational claim. If the interactive toolkit group does not retain more or learn faster, the paper's core justification for the tool's design is contradicted.","tokens_in":7304,"feed_emoji":"🧠","tokens_out":7586,"duration_ms":71019,"temperature":0.7,"pith_summary":"asanAI is an open-source, browser-based toolkit whose stated aim is to remove every conventional barrier to machine learning: no programming, no installation, no server, and no persistent internet connection. A user assembles a neural network by adding layers in a graphical panel, watches training in live plots and color-coded weight updates, and can export the finished model as Python code for use elsewhere. The intended benefit is both educational and practical: teachers can run ML lessons on whatever devices a classroom already has, and researchers can draft and test ideas without setting up an environment. The paper's claim is that this lowers the entry barrier for anyone, regardless of technical expertise or resources.","feed_headline":"No code, no install, no internet: ML training moves into the browser","feed_subtitle":"A free, offline-first toolkit lets non-programmers design, train, and test neural networks on any device.","key_machinery":"The carrying mechanism is the browser itself as a complete runtime. The toolkit keeps one neural-network model as a singleton in memory, backed by a JavaScript deep-learning library, and recompiles the model on every interface change while preserving weights when shapes allow; a 200-millisecond validation loop checks every field and blocks training until the configuration is valid. On top of that core sit the layer panel, the functional grouping of layers into feature detection and classification, and three user modes (Introductory, Beginner, Expert) that adjust how much guidance the interface gives. This architecture is what makes the no-code, offline, and export-oriented promises hang together.","core_discovery":"The central claim is that a no-code, offline-first, browser-based toolkit can be a legitimate on-ramp to machine learning for users from first-time students to working researchers, without cutting off the path to professional deployment. asanAI's design pairs a visual, layer-by-layer model builder with immediate feedback and multiple views of what the network is doing: full-connectivity graphs, filter-style layer views, feature maps, GradCAM heatmaps, and a math mode that shows the exact equations for simple networks. The paper presents the toolkit as the demonstration of its own thesis: it trains and evaluates neural networks entirely on the local device, runs on smartphones, continues to work after the page is loaded offline, and exports models in a form that can run in standard Python environments. The authors report that the toolkit is already in use by researchers, trainers, and teachers, which is the evidence offered for its practical reach.","pith_inferences":["The strongest untested extension is longitudinal transfer: whether learners who build models visually are later able to read or modify the exported Python code; the toolkit makes that experiment possible because code generation is built in, but the paper does not report such a measurement.","The same feature-map and GradCAM views that are framed as teaching aids could be repurposed as quick bias-detection tools for classroom image datasets, showing which image regions drive a classification.","Because browser-based training caps model and dataset size, the realistic niche suggested by the design is small- to medium-sized prototyping and teaching; a natural test is whether the Python export path lets users scale up once they hit the browser's limits."],"forward_implications":["Classrooms without reliable internet can still teach modern ML topics, since every device with a modern browser becomes a potential training station.","A user can go from a visual scratch model to a standalone Python program by exporting the auto-generated code, so the no-code environment can serve as a bridge rather than a dead end.","Sensitive data never has to leave the device, which makes the toolkit usable in schools or institutions that cannot upload data to a cloud service.","The MIT-style license lets schools, researchers, and hobbyists modify the tool, so the toolkit can be adapted to specific curricula or workflows."],"supporting_citations":[{"why":"Supplies the in-browser deep-learning engine that makes local training and model export possible.","marker":"Smilkov et al. 2019"},{"why":"The server-dependent no-code classification tool whose limitations motivate asanAI's offline-first design.","marker":"Carney et al. 2020"},{"why":"The OS-restricted no-code tool that motivates cross-device browser delivery.","marker":"Microsoft 2021"},{"why":"The web ML library that still requires programming skills, marking the gap asanAI targets.","marker":"Shiffman 2020"},{"why":"Cited for the educational advantages of browser-based ML toolkits, underpinning the learning rationale.","marker":"Lane 2024"},{"why":"Provides the input-maximization method used for feature-map visualizations.","marker":"Zeiler and Fergus 2013"},{"why":"Provides the gradient-localization method used for GradCAM heatmaps.","marker":"Selvaraju et al. 2016"}],"fun_headline_variants":["ML toolkit runs fully in browser, no code or install needed","Train neural nets in your browser, offline, with zero coding","No-code ML: design and train models right in your tab","Offline-first ML: no install, no server, just your browser","Browser-based ML: build and train models without writing code"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that interacting with immediate visual feedback produces higher engagement and better information retention, which the design section asserts without a user study or citation; if that premise fails, the toolkit's central value as an educational tool is substantially weakened.","fun_headline_variants_meta":{"raw":{"variants":["ML toolkit runs fully in browser, no code or install needed","Train neural nets in your browser, offline, with zero coding","No-code ML: design and train models right in your tab","Offline-first ML: no install, no server, just your browser","Browser-based ML: build and train models without writing code"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000261,"raw_usage":{"total_tokens":1615,"prompt_tokens":988,"completion_tokens":627,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":540}},"tokens_in":604,"tokens_out":627,"duration_ms":5758,"temperature":1.0,"reasoning_tokens":540,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:46:52.203152+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled experiment with two matched groups of learners—one building and training networks in asanAI, one receiving the same concepts through a static slide-based lesson—with a delayed test of concept understanding would settle the educational claim. If the interactive toolkit group does not retain more or learn faster, the paper's core justification for the tool's design is contradicted.","supporting_citations":[],"review_version":1}