Haoran Zhao

Haoran Zhao, Jake Ryland Williams

While Large Language Models (LLMs) become ever more dominant, classic pre-trained word embeddings sustain their relevance through computational efficiency and nuanced linguistic interpretation. Drawing from recent studies demonstrating that the convergence of GloVe and word2vec optimizations all tend towards log-co-occurrence matrix variants, we construct a novel word representation system called Bit-cipher that eliminates the need of backpropagation while leveraging contextual information and hyper-efficient dimensionality reduction techniques based on unigram frequency, providing strong interpretability, alongside efficiency. We use the bit-cipher algorithm to train word vectors via a two-step process that critically relies on a hyperparameter -- bits -- that controls the vector dimension. While the first step trains the bit-cipher, the second utilizes it under two different aggregation modes -- summation or concatenation -- to produce contextually rich representations from word co-occurrences. We extend our investigation into bit-cipher's efficacy, performing probing experiments on part-of-speech (POS) tagging and named entity recognition (NER) to assess its competitiveness with classic embeddings like word2vec and GloVe. Additionally, we explore its applicability in LM training and fine-tuning. By replacing embedding layers with cipher embeddings, our experiments illustrate the notable efficiency of cipher in accelerating the training process and attaining better optima compared to conventional training paradigms. Experiments on the integration of bit-cipher embedding layers with Roberta, T5, and OPT, prior to or as a substitute for fine-tuning, showcase a promising enhancement to transfer learning, allowing rapid model convergence while preserving competitive performance.

Jake Ryland Williams, Haoran Zhao

Iterative approximation methods using backpropagation enable the optimization of neural networks, but they remain computationally expensive, especially when used at scale. This paper presents an efficient alternative for optimizing neural networks that reduces the costs of scaling neural networks and provides high-efficiency optimizations for low-resource applications. We will discuss a general result about feed-forward neural networks and then extend this solution to compositional (mult-layer) networks, which are applied to a simplified transformer block containing feed-forward and self-attention layers. These models are used to train highly-specified and complex multi-layer neural architectures that we refer to as self-attentive feed-forward unit (SAFFU) layers, which we use to develop a transformer that appears to generalize well over small, cognitively-feasible, volumes of data. Testing demonstrates explicit solutions outperform models optimized by backpropagation alone. Moreover, further application of backpropagation after explicit solutions leads to better optima from smaller scales of data, training effective models from much less data is enabled by explicit solution warm starts. We then carry out ablation experiments training a roadmap of about 250 transformer models over 1-million tokens to determine ideal settings. We find that multiple different architectural variants produce highly-performant models, and discover from this ablation that some of the best are not the most parameterized. This appears to indicate well-generalized models could be reached using less data by using explicit solutions, and that architectural exploration using explicit solutions pays dividends in guiding the search for efficient variants with fewer parameters, and which could be incorporated into low-resource hardware where AI might be embodied.

Xiangyu Dong, Haoran Zhao, Jiang Gao et al.

Recent advances in vision-language navigation (VLN) were mainly attributed to emerging large language models (LLMs). These methods exhibited excellent generalization capabilities in instruction understanding and task reasoning. However, they were constrained by the fixed knowledge bases and reasoning abilities of LLMs, preventing fully incorporating experiential knowledge and thus resulting in a lack of efficient evolutionary capacity. To address this, we drew inspiration from the evolution capabilities of natural agents, and proposed a self-evolving VLN framework (SE-VLN) to endow VLN agents with the ability to continuously evolve during testing. To the best of our knowledge, it was the first time that an multimodal LLM-powered self-evolving VLN framework was proposed. Specifically, SE-VLN comprised three core modules, i.e., a hierarchical memory module to transfer successful and failure cases into reusable knowledge, a retrieval-augmented thought-based reasoning module to retrieve experience and enable multi-step decision-making, and a reflection module to realize continual evolution. Comprehensive tests illustrated that the SE-VLN achieved navigation success rates of 57% and 35.2% in unseen environments, representing absolute performance improvements of 23.9% and 15.0% over current state-of-the-art methods on R2R and REVERSE datasets, respectively. Moreover, the SE-VLN showed performance improvement with increasing experience repository, elucidating its great potential as a self-evolving agent framework for VLN.

Haoran Zhao, Wayne Isaac Tan Uy

Stochastic generators are essential to produce synthetic realizations that preserve target statistical properties. We propose GenFormer, a stochastic generator for spatio-temporal multivariate stochastic processes. It is constructed using a Transformer-based deep learning model that learns a mapping between a Markov state sequence and time series values. The synthetic data generated by the GenFormer model preserves the target marginal distributions and approximately captures other desired statistical properties even in challenging applications involving a large number of spatial locations and a long simulation horizon. The GenFormer model is applied to simulate synthetic wind speed data at various stations in Florida to calculate exceedance probabilities for risk management.

Haoran Zhao, Xin Sun, Junyu Dong et al.

Recently, distillation approaches are suggested to extract general knowledge from a teacher network to guide a student network. Most of the existing methods transfer knowledge from the teacher network to the student via feeding the sequence of random mini-batches sampled uniformly from the data. Instead, we argue that the compact student network should be guided gradually using samples ordered in a meaningful sequence. Thus, it can bridge the gap of feature representation between the teacher and student network step by step. In this work, we provide a curriculum learning knowledge distillation framework via instance-level sequence learning. It employs the student network of the early epoch as a snapshot to create a curriculum for the student network's next training phase. We carry out extensive experiments on CIFAR-10, CIFAR-100, SVHN and CINIC-10 datasets. Compared with several state-of-the-art methods, our framework achieves the best performance with fewer iterations.

Haoran Zhao, Xin Sun, Junyu Dong et al.

Knowledge distillation has been widely used to produce portable and efficient neural networks which can be well applied on edge devices for computer vision tasks. However, almost all top-performing knowledge distillation methods need to access the original training data, which usually has a huge size and is often unavailable. To tackle this problem, we propose a novel data-free approach in this paper, named Dual Discriminator Adversarial Distillation (DDAD) to distill a neural network without any training data or meta-data. To be specific, we use a generator to create samples through dual discriminator adversarial distillation, which mimics the original training data. The generator not only uses the pre-trained teacher's intrinsic statistics in existing batch normalization layers but also obtains the maximum discrepancy from the student model. Then the generated samples are used to train the compact student network under the supervision of the teacher. The proposed method obtains an efficient student network which closely approximates its teacher network, despite using no original training data. Extensive experiments are conducted to to demonstrate the effectiveness of the proposed approach on CIFAR-10, CIFAR-100 and Caltech101 datasets for classification tasks. Moreover, we extend our method to semantic segmentation tasks on several public datasets such as CamVid and NYUv2. All experiments show that our method outperforms all baselines for data-free knowledge distillation.

Haoran Zhao, Kun Gong, Xin Sun et al.

Knowledge distillation is a popular paradigm for learning portable neural networks by transferring the knowledge from a large model into a smaller one. Most existing approaches enhance the student model by utilizing the similarity information between the categories of instance level provided by the teacher model. However, these works ignore the similarity correlation between different instances that plays an important role in confidence prediction. To tackle this issue, we propose a novel method in this paper, called similarity transfer for knowledge distillation (STKD), which aims to fully utilize the similarities between categories of multiple samples. Furthermore, we propose to better capture the similarity correlation between different instances by the mixup technique, which creates virtual samples by a weighted linear interpolation. Note that, our distillation loss can fully utilize the incorrect classes similarities by the mixed labels. The proposed approach promotes the performance of student model as the virtual sample created by multiple images produces a similar probability distribution in the teacher and student networks. Experiments and ablation studies on several public classification datasets including CIFAR-10,CIFAR-100,CINIC-10 and Tiny-ImageNet verify that this light-weight method can effectively boost the performance of the compact student model. It shows that STKD substantially has outperformed the vanilla knowledge distillation and has achieved superior accuracy over the state-of-the-art knowledge distillation methods.