Exploring the Knowledge Mismatch Hypothesis: Hallucination Propensity in Small Models Fine-tuned on Data from Larger Models

Recently, there has been an explosion of large language models created through fine-tuning with data from larger models. These small models able to produce outputs that appear qualitatively similar to significantly larger models. However, one of the key limitations that have been observed with these models is their propensity to hallucinate significantly more often than larger models. In particular, they have been observed to generate coherent outputs that involve factually incorrect information and spread misinformation, toxicity, and stereotypes. There are many potential causes of hallucination, of which, one hypothesis is that fine-tuning a model on data produced by a larger model leads to a knowledge mismatch which contributes to hallucination. In particular, it is hypothesized that there is a mismatch between the knowledge that is fed to the model to fine-tune it and the knowledge that is already present in the graph. Fine-tuning the model on data that has such mismatch could contribute to an increased propensity to hallucinate. We show that on an unseen test set, a smaller model fine-tuned on data generated from a larger model produced more wrong answers when compared to models fine-tuned on data created by the small model, which confirms the hypothesis.

Combining Neural Architecture Search and Automatic Code Optimization: A Survey

Deep Learning models have experienced exponential growth in complexity and resource demands in recent years. Accelerating these models for efficient execution on resource-constrained devices has become more crucial than ever. Two notable techniques employed to achieve this goal are Hardware-aware Neural Architecture Search (HW-NAS) and Automatic Code Optimization (ACO). HW-NAS automatically designs accurate yet hardware-friendly neural networks, while ACO involves searching for the best compiler optimizations to apply on neural networks for efficient mapping and inference on the target hardware. This survey explores recent works that combine these two techniques within a single framework. We present the fundamental principles of both domains and demonstrate their sub-optimality when performed independently. We then investigate their integration into a joint optimization process that we call Hardware Aware-Neural Architecture and Compiler Optimizations co-Search (NACOS).

Curriculum Learning for Small Code Language Models

Code language models have emerged as useful tools for various programming tasks, yet they often struggle when it comes to complex ones. In this paper, we explore the potential of curriculum learning in enhancing the performance of these models. While prior research has suggested that curriculum learning does not necessarily help in improving the performance of language models, our results surprisingly show that this may not be the case for code language models. We demonstrate that a well-designed curriculum learning approach significantly improves the accuracy of small decoder-only code language models on the task of code execution, while its effect on code completion is less significant. To explore the potential of curriculum learning, we train multiple GPT models with 1 million parameters each to predict the next token and evaluate them on code completion and execution tasks. Our contributions include proposing a novel code difficulty assessment metric by combining software code measures, investigating the effectiveness of Curriculum Learning for code language models, and introducing a Novel Curriculum Learning schedule that enhances the performance of small decoder-only language models in code execution tasks. The results of this paper open the door for more research on the use of curriculum learning for code language models.

LOOPer: A Learned Automatic Code Optimizer For Polyhedral Compilers

While polyhedral compilers have shown success in implementing advanced code transformations, they still have challenges in selecting the most profitable transformations that lead to the best speedups. This has motivated the use of machine learning to build cost models to guide the search for polyhedral optimizations. State-of-the-art polyhedral compilers have demonstrated a viable proof-of-concept of this approach. While such a proof-of-concept has shown promise, it still has significant limitations. State-of-the-art polyhedral compilers that use a deep-learning cost model only support a small subset of affine transformations, limiting their ability to apply complex code transformations. They also only support simple programs that have a single loop nest and a rectangular iteration domain, limiting their applicability to many programs. These limitations significantly impact the generality of such compilers and autoschedulers and put into question the whole approach. In this paper, we introduce LOOPer, the first polyhedral autoscheduler that uses a deep-learning based cost model and covers a large set of affine transformations and programs. It supports the exploration of a large set of affine transformations, allowing the application of complex sequences of polyhedral transformations. It also supports the optimization of programs with multiple loop nests and with rectangular and non-rectangular iteration domains, allowing the optimization of an extensive set of programs. We implement and evaluate LOOPer and show that it achieves speedups over the state-of-the-art. On the Polybench benchmark, LOOPer achieves a geometric mean speedup of 1.59x over Tiramisu. LOOPer also achieves competitive speedups with a geometric mean speedup of 1.34x over Pluto, a state-of-the-art polyhedral compiler that does not use a machine-learning based cost model.

Leveraging High-Resolution Features for Improved Deep Hashing-based Image Retrieval

Deep hashing techniques have emerged as the predominant approach for efficient image retrieval. Traditionally, these methods utilize pre-trained convolutional neural networks (CNNs) such as AlexNet and VGG-16 as feature extractors. However, the increasing complexity of datasets poses challenges for these backbone architectures in capturing meaningful features essential for effective image retrieval. In this study, we explore the efficacy of employing high-resolution features learned through state-of-the-art techniques for image retrieval tasks. Specifically, we propose a novel methodology that utilizes High-Resolution Networks (HRNets) as the backbone for the deep hashing task, termed High-Resolution Hashing Network (HHNet). Our approach demonstrates superior performance compared to existing methods across all tested benchmark datasets, including CIFAR-10, NUS-WIDE, MS COCO, and ImageNet. This performance improvement is more pronounced for complex datasets, which highlights the need to learn high-resolution features for intricate image retrieval tasks. Furthermore, we conduct a comprehensive analysis of different HRNet configurations and provide insights into the optimal architecture for the deep hashing task

Automatic Generation of Python Programs Using Context-Free Grammars

In recent years, data has emerged as the new gold, serving as a powerful tool for creating intelligent systems. However, procuring high-quality data remains challenging, especially for code. To address this, we developed TinyPy Generator, a tool that generates random Python programs using a context-free grammar. The generated programs are guaranteed to be correct by construction. Our system uses custom production rules (in the Backus-Naur Form (BNF) format) to recursively generate code. This allows us to generate code with different levels of complexity, ranging from code containing only assignments to more complex code containing conditionals and loops. Our proposed tool enables effortless large-scale Python code generation, beneficial for a wide range of applications. TinyPy Generator is particularly useful in the field of machine learning, where it can generate substantial amounts of Python code for training Python language models. Additionally, researchers who are studying programming languages can utilize this tool to create datasets for their experiments, which can help validate the robustness of code interpreters or compilers. Unlike existing research, we have open-sourced our implementation. This allows customization according to user needs and extends potential usage to other languages.

Progress Report: A Deep Learning Guided Exploration of Affine Unimodular Loop Transformations

In this paper, we present a work in progress about a deep learning based approach for automatic code optimization in polyhedral compilers. The proposed technique explores combinations of affine and non-affine loop transformations to find the sequence of transformations that minimizes the execution time of a given program. This exploration is guided by a deep learning based cost model that evaluates the speedup that each sequence of transformations would yield. Preliminary results show that the proposed techniques achieve a 2.35x geometric mean speedup over state of the art polyhedral compilers (Pluto).

A Deep Learning Based Cost Model for Automatic Code Optimization

Enabling compilers to automatically optimize code has been a longstanding goal for the compiler community. Efficiently solving this problem requires using precise cost models. These models predict whether applying a sequence of code transformations reduces the execution time of the program. Building an analytical cost model to do so is hard in modern x86 architectures due to the complexity of the microarchitecture. In this paper, we present a novel deep learning based cost model for automatic code optimization. This model was integrated in a search method and implemented in the Tiramisu compiler to select the best code transformations. The input of the proposed model is a set of simple features representing the unoptimized code and a sequence of code transformations. The model predicts the speedup expected when the code transformations are applied. Unlike previous models, the proposed one works on full programs and does not rely on any heavy feature engineering. The proposed model has only 16% of mean absolute percentage error in predicting speedups on full programs. The proposed model enables Tiramisu to automatically find code transformations that match or are better than state-of-the-art compilers without requiring the same level of heavy feature engineering required by those compilers.

Hardware Acceleration of Sparse and Irregular Tensor Computations of ML Models: A Survey and Insights

Machine learning (ML) models are widely used in many domains including media processing and generation, computer vision, medical diagnosis, embedded systems, high-performance and scientific computing, and recommendation systems. For efficiently processing these computational- and memory-intensive applications, tensors of these over-parameterized models are compressed by leveraging sparsity, size reduction, and quantization of tensors. Unstructured sparsity and tensors with varying dimensions yield irregular-shaped computation, communication, and memory access patterns; processing them on hardware accelerators in a conventional manner does not inherently leverage acceleration opportunities. This paper provides a comprehensive survey on how to efficiently execute sparse and irregular tensor computations of ML models on hardware accelerators. In particular, it discusses additional enhancement modules in architecture design and software support; categorizes different hardware designs and acceleration techniques and analyzes them in terms of hardware and execution costs; highlights further opportunities in terms of hardware/software/algorithm co-design optimizations and joint optimizations among described hardware and software enhancement modules. The takeaways from this paper include: understanding the key challenges in accelerating sparse, irregular-shaped, and quantized tensors; understanding enhancements in acceleration systems for supporting their efficient computations; analyzing trade-offs in opting for a specific type of design enhancement; understanding how to map and compile models with sparse tensors on the accelerators; understanding recent design trends for efficient accelerations and further opportunities.

TIRAMISU: A Polyhedral Compiler for Dense and Sparse Deep Learning

In this paper, we demonstrate a compiler that can optimize sparse and recurrent neural networks, both of which are currently outside of the scope of existing neural network compilers (sparse neural networks here stand for networks that can be accelerated with sparse tensor algebra techniques). Our demonstration includes a mapping of sparse and recurrent neural networks to the polyhedral model along with an implementation of our approach in TIRAMISU, our state-of-the-art polyhedral compiler. We evaluate our approach on a set of deep learning benchmarks and compare our results with hand-optimized industrial libraries. Our results show that our approach at least matches Intel MKL-DNN and in some cases outperforms it by 5x (on multicore-CPUs).