Opening the AI black box: program synthesis via mechanistic interpretability

• URL: http://arxiv.org/abs/2402.05110v1
• Date: Wed, 7 Feb 2024 18:59:12 GMT
• Title: Opening the AI black box: program synthesis via mechanistic interpretability
• Authors: Eric J. Michaud, Isaac Liao, Vedang Lad, Ziming Liu, Anish Mudide, Chloe Loughridge, Zifan Carl Guo, Tara Rezaei Kheirkhah, Mateja Vukeli\'c, Max Tegmark
• Abstract summary: We present a novel method for program synthesis based on automated mechanistic interpretability of neural networks trained to perform the desired task, auto-distilling the learned algorithm into Python code. We test MIPS on a benchmark of 62 algorithmic tasks that can be learned by an RNN and find it highly complementary to GPT-4. As opposed to large language models, this program synthesis technique makes no use of (and is therefore not limited by) human training data such as algorithms and code from GitHub.
• Score: 12.849101734204456
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: We present MIPS, a novel method for program synthesis based on automated mechanistic interpretability of neural networks trained to perform the desired task, auto-distilling the learned algorithm into Python code. We test MIPS on a benchmark of 62 algorithmic tasks that can be learned by an RNN and find it highly complementary to GPT-4: MIPS solves 32 of them, including 13 that are not solved by GPT-4 (which also solves 30). MIPS uses an integer autoencoder to convert the RNN into a finite state machine, then applies Boolean or integer symbolic regression to capture the learned algorithm. As opposed to large language models, this program synthesis technique makes no use of (and is therefore not limited by) human training data such as algorithms and code from GitHub. We discuss opportunities and challenges for scaling up this approach to make machine-learned models more interpretable and trustworthy.

Related papers

• MARIO: MAth Reasoning with code Interpreter Output -- A Reproducible Pipeline [12.186691561822256]
  We postulate that the inherent nature of large language models (LLMs) presents challenges in modeling mathematical reasoning. This paper introduces a novel math dataset, enhanced with a capability to utilize a Python code interpreter. We propose a tentative, easily replicable protocol for the fine-tuning of math-specific LLMs.
  arXiv  Detail & Related papers  (2024-01-16T08:08:01Z)
• On Statistical Learning of Branch and Bound for Vehicle Routing Optimization [3.6922704509753084]
  We train neural networks to emulate the decision-making process of the computationally expensive Strong Branching strategy. We find that this approach can match or improve upon the performance of the branch and bound algorithm.
  arXiv  Detail & Related papers  (2023-10-15T23:59:57Z)
• Guess & Sketch: Language Model Guided Transpilation [59.02147255276078]
  Learned transpilation offers an alternative to manual re-writing and engineering efforts. Probabilistic neural language models (LMs) produce plausible outputs for every input, but do so at the cost of guaranteed correctness. Guess & Sketch extracts alignment and confidence information from features of the LM then passes it to a symbolic solver to resolve semantic equivalence.
  arXiv  Detail & Related papers  (2023-09-25T15:42:18Z)
• The Clock and the Pizza: Two Stories in Mechanistic Explanation of Neural Networks [59.26515696183751]
  We show that algorithm discovery in neural networks is sometimes more complex. We show that even simple learning problems can admit a surprising diversity of solutions.
  arXiv  Detail & Related papers  (2023-06-30T17:59:13Z)
• Stealing the Decoding Algorithms of Language Models [56.369946232765656]
  A key component of generating text from modern language models (LM) is the selection and tuning of decoding algorithms. In this work, we show, for the first time, that an adversary with typical API access to an LM can steal the type and hyper parameters of its decoding algorithms. Our attack is effective against popular LMs used in text generation APIs, including GPT-2, GPT-3 and GPT-Neo.
  arXiv  Detail & Related papers  (2023-03-08T17:15:58Z)
• NP4G : Network Programming for Generalization [0.0]
  We propose NP4G: Network Programming for Generalization, which can automatically generate programs by inductive inference. As an example, we show the bitwise NOT operation programs are acquired in a comparatively short time and at a rate of about 7 in 10 running.
  arXiv  Detail & Related papers  (2022-12-08T06:18:44Z)
• CGEMs: A Metric Model for Automatic Code Generation using GPT-3 [0.0]
  This work aims to validate AI-generated content using theoretical proofs or by using Monte-Carlo simulation methods. In this case, we use the latter approach to test/validate a statistically significant number of samples. The various metrics that are garnered in this work to support the evaluation of generated code are as follows: Compilation, NL description to logic conversion, number of edits needed, some of the commonly used static-code metrics and NLP metrics.
  arXiv  Detail & Related papers  (2021-08-23T13:28:57Z)
• Using Machine Learning for Particle Track Identification in the CLAS12 Detector [0.0]
  This article describes the development of four machine learning (ML) models that assist the tracking by identifying valid track candidates. An algorithm with recommended track candidates was implemented as part of the CLAS12 reconstruction software. The resulting software achieved accuracy of greater than 99% and resulted in an end-to-end speedup of 35% compared to existing algorithms.
  arXiv  Detail & Related papers  (2020-08-28T21:57:38Z)
• Strong Generalization and Efficiency in Neural Programs [69.18742158883869]
  We study the problem of learning efficient algorithms that strongly generalize in the framework of neural program induction. By carefully designing the input / output interfaces of the neural model and through imitation, we are able to learn models that produce correct results for arbitrary input sizes.
  arXiv  Detail & Related papers  (2020-07-07T17:03:02Z)
• Hybrid Models for Learning to Branch [81.93868699246214]
  We propose a new hybrid architecture for efficient branching on CPU machines. The proposed architecture combines the expressive power of GNNs with computationally inexpensive multi-layer perceptrons (MLP) for branching.
  arXiv  Detail & Related papers  (2020-06-26T21:03:45Z)
• Channel Assignment in Uplink Wireless Communication using Machine Learning Approach [54.012791474906514]
  This letter investigates a channel assignment problem in uplink wireless communication systems. Our goal is to maximize the sum rate of all users subject to integer channel assignment constraints. Due to high computational complexity, machine learning approaches are employed to obtain computational efficient solutions.
  arXiv  Detail & Related papers  (2020-01-12T15:54:20Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.