SnAp approximates RTRL for online training of sparse recurrent networks.
problem Training large sparse recurrent networks online is computationally expensive.
method Sparse n-step Approximation (SnAp) of the RTRL influence matrix.
result SnAp with n=2 remains tractable for highly sparse networks and outperforms backpropagation through time.
VSML unifies meta learning concepts and enables simple backpropagation.
problem Improving and unifying meta learning concepts for neural networks.
method Unified approach using variable shared meta learning and simple weight-sharing.
result Simple backpropagation can be implemented and meta learned without gradient calculation.
New method reduces memory usage in deep HRNNs by replacing gradient backpropagation with local losses.
problem Memory constraints in training deep hierarchical RNNs.
method Replace gradient backpropagation with locally computable losses in deep HRNNs.
result Memory requirements reduced by a factor exponential in hierarchy depth.
A new algorithm reduces the computational cost of RTRL while maintaining performance.
problem Large computational costs in RTRL for learning long-term dependencies.
method Optimal Kronecker-Sum Approximation (OK) of RTRL.
result OK is optimal for a class of RTRL approximations and matches TBPTT in performance.
Despite recent advances in training recurrent neural networks (RNNs), capturing long-term dependencies in sequences remains a fundamental challenge. Most approaches use backpropagation through time (BPTT), which is difficult to scale to very long sequences. This paper proposes a simple method that improves the ability …
New learning rules for wide neural networks without backpropagation.
problem Training wide neural networks efficiently and without backpropagation.
method Input-weight alignment driven by gradient descent in the NTK regime.
result Biologically-motivated learning rules equivalent to backpropagation in wide networks.
Reversible RNNs reduce memory usage in training without sacrificing performance.
problem Memory-intensive training of RNNs limits model flexibility.
method Developed a scheme for perfect reversible RNNs with forgetting, reducing memory by 10-15x.
result Achieved comparable performance to traditional models with reduced activation memory cost.
This note presents in a technical though hopefully pedagogical way the three most common forms of neural network architectures: Feedforward, Convolutional and Recurrent. For each network, their fundamental building blocks are detailed. The forward pass and the update rules for the backpropagation algorithm are then der…
Adaptive TBPTT controls gradient bias in RNNs for faster convergence.
problem Choosing optimal truncation length in TBPTT for RNNs is difficult.
method Adaptive TBPTT converts lag selection to bias control, estimating optimal truncation length during training.
result Adaptive TBPTT improves convergence rate and computational efficiency in RNNs.
A major drawback of backpropagation through time (BPTT) is the difficulty of learning long-term dependencies, coming from having to propagate credit information backwards through every single step of the forward computation. This makes BPTT both computationally impractical and biologically implausible. For this reason,…
ERNN improves RNN accuracy and stability with time-delayed self-feedback.
problem Inaccuracy and instability in RNNs.
method Augmenting RNN with a time-delayed self-feedback loop to stabilize hidden state transitions.
result ERNN achieves state-of-the-art results on benchmark datasets.
A new deep approach to Kalman filtering integrates uncertainty estimates efficiently.
problem Integrating uncertainty estimates into deep time-series models.
method Proposes a Recurrent Kalman Network (RKN) that learns directly using backpropagation.
result RKN obtains more accurate uncertainty estimates and slightly improved prediction performance.
We introduce the "NoBackTrack" algorithm to train the parameters of dynamical systems such as recurrent neural networks. This algorithm works in an online, memoryless setting, thus requiring no backpropagation through time, and is scalable, avoiding the large computational and memory cost of maintaining the full gradie…
End-to-end training of recurrent neural networks for biological sequences.
problem Training recurrent neural networks for biological sequence classification.
method Generalized convolutional kernel networks with recurrent structures.
result Outperforms existing methods for protein classification tasks.
RTRL optimizes long sequences without truncation, converging to loss minima.
problem Inaccuracies in TBPTT for long sequences.
method Online optimization with exact gradient calculation.
result RTRL converges to loss minima for a class of RNNs.
A new model uses 'ghost units' to enable efficient backpropagation in deep neural networks.
problem How to achieve efficient backpropagation in deep neural networks with biological plausibility.
method Introduces 'ghost units' to cancel feedback, enabling efficient error backpropagation.
result Demonstrates that the model can approximate error gradients and achieve good performance on classification tasks.
RNNs are crucial for text and speech tasks, explained in this overview.
problem Understanding RNNs for advanced AI tasks.
method Overview of RNN concepts and recent advances.
result Fundamentals and recent techniques in RNNs.
SparseProp speeds up SNN simulations and training by four orders of magnitude.
problem Efficiently simulating and training large spiking neural networks.
method Event-based algorithm that reduces computational cost from O(N) to O(log(N)) per spike.
result Numerically exact simulations of large spiking networks and efficient training using backpropagation.
How can we build agents that keep learning from experience, quickly and efficiently, after their initial training? Here we take inspiration from the main mechanism of learning in biological brains: synaptic plasticity, carefully tuned by evolution to produce efficient lifelong learning. We show that plasticity, just li…
In this work we explore a straightforward variational Bayes scheme for Recurrent Neural Networks. Firstly, we show that a simple adaptation of truncated backpropagation through time can yield good quality uncertainty estimates and superior regularisation at only a small extra computational cost during training, also re…
We parallelize backpropagation for deep learning models, achieving significant speedups.
problem Sequential dependency in backpropagation limits scalability on parallel systems.
method Reformulated backpropagation as a scan operation, using Blelloch scan algorithm.
result Up to 2.75x speedup on overall training time and 108x on backward pass.
Recurrent neural networks are nowadays successfully used in an abundance of applications, going from text, speech and image processing to recommender systems. Backpropagation through time is the algorithm that is commonly used to train these networks on specific tasks. Many deep learning frameworks have their own imple…
Despite all the impressive advances of recurrent neural networks, sequential data is still in need of better modelling. Truncated backpropagation through time (TBPTT), the learning algorithm most widely used in practice, suffers from the truncation bias, which drastically limits its ability to learn long-term dependenc…
Paper proposes an ELM-trained recurrent BBFNN for faster and more robust time series prediction.
problem Training recurrent BBFNNs is computationally intensive and prone to noise.
method Adopted an ELM for quick training, added a recurrent structure to BBFNN.
result The proposed network outperforms traditional networks in accuracy and robustness.
ERNNs evolve hidden states on an ODE's equilibrium manifold to mitigate vanishing and exploding gradients.
problem Vanishing and exploding gradients in RNNs.
method Develop a novel family of RNNs (ERNNs) that evolve hidden states on the equilibrium manifold of an ODE.
result ERNNs achieve state-of-the-art accuracy with 3-10x speedups and 1.5-3x model size reduction.
New FPI layers enable efficient backpropagation in deep networks.
problem Designing deep neural networks to handle complex constraints.
method Fixed-point iteration layers for forward and backward propagation.
result Backward FPI layer simplifies gradient calculation without explicit Jacobian.
Reinforcement learning (RL) agents performing complex tasks must be able to remember observations and actions across sizable time intervals. This is especially true during the initial learning stages, when exploratory behaviour can increase the delay between specific actions and their effects. Many new or popular appro…
Single-layer Transformer can approximate any sequence mapping.
problem Lack of theoretical understanding of Transformers.
method Review of linear algebra, probability, and optimization concepts; detailed analysis of Transformer architecture.
result A single-layer Transformer can approximate any continuous sequence-to-sequence mapping to arbitrary precision.
New neural network theory mimics physics laws, making computations more plausible.
problem Neural networks lack biological plausibility in computations.
method Variational framework of the least action principle, local in space and time.
result SpatioTemporal Local Propagation (STLP) scheme is biologically plausible.
Bayesian RNNs improve safety in RL tasks.
problem Lack of explainability and safety in RL algorithms.
method Fully Bayesian Recurrent Neural Networks (RNNs) with uncertainty estimates.
result Significantly outperforms existing methods in collision avoidance tasks.
Time series are widely used as signals in many classification/regression tasks. It is ubiquitous that time series contains many missing values. Given multiple correlated time series data, how to fill in missing values and to predict their class labels? Existing imputation methods often impose strong assumptions of the …
EP matches BPTT gradients in discrete-time RNNs, improving training efficiency.
problem Training convergent RNNs efficiently with EP.
method Introduced a discrete-time version of EP with simplified equations and backward gradients.
result EP's neural and weight updates are step-by-step equal to BPTT's, with gradients computed backward.
Gradient flossing stabilizes RNN training by controlling Lyapunov exponents.
problem Gradient instability in RNNs leading to exploding and vanishing gradients.
method Regularizing Lyapunov exponents through backpropagation using differentiable linear algebra.
result Gradient flossing improves RNN training success rate and convergence speed.
The paper improves DFA for CNN and RNN training to match BP accuracy.
problem Low accuracy in CNN and RNN training using DFA.
method Divide network into modules, apply DFA within, use sparse backward weight, and incorporate dilated convolution and sparse matrix multiplication.
result Achieves BP-level accuracy in CNN and RNN training.
Advancements in parallel processing have lead to a surge in multilayer perceptrons' (MLP) applications and deep learning in the past decades. Recurrent Neural Networks (RNNs) give additional representational power to feedforward MLPs by providing a way to treat sequential data. However, RNNs are hard to train using con…
Belief propagation recovers backpropagation results.
problem Connection between backpropagation and belief propagation poorly understood.
method Converted backpropagation input to belief propagation input and showed results.
result Backpropagation is a special case of belief propagation.
Tensor programs prove neural network limits for any architecture.
problem Understanding the limits of neural networks of any architecture.
method Prove convergence of neural network's Tangent Kernel (NTK) to a deterministic limit as network widths increase.
result Identify conditions for correct NTK limit calculation based on gradient independence assumption.
Oja's rule improves neural network training without engineered tricks.
problem Training deep neural networks with biological constraints.
method Incorporating Oja's plasticity rule into error-driven training.
result Stable, efficient learning in feedforward and recurrent architectures.
In a physical neural system, where storage and processing are intimately intertwined, the rules for adjusting the synaptic weights can only depend on variables that are available locally, such as the activity of the pre- and post-synaptic neurons, resulting in local learning rules. A systematic framework for studying t…
Paper proposes DAG-DB for learning discrete DAGs via backpropagation.
problem Learning Directed Acyclic Graphs (DAGs) from data.
method DAG-DB uses Discrete Backpropagation with I-MLE and Straight-Through Estimation.
result DAG-DB learns DAGs effectively using probabilistic sampling and backpropagation.
This paper presents a new approach for training artificial neural networks using techniques for solving the constraint satisfaction problem (CSP). The quotient gradient system (QGS) is a trajectory-based method for solving the CSP. This study converts the training set of a neural network into a CSP and uses the QGS to …
Analyzes learning dynamics of RNNs under locality constraints.
problem Understanding learning dynamics in RNNs with locality constraints.
method Dynamical systems theory applied to data-aligned linear RNNs.
result RFLO solutions are restricted to low-rank perturbations of initial parameters.
GAIT-prop derives a biologically plausible learning rule from backpropagation.
problem Biological implausibility in traditional backpropagation for neural networks.
method GAIT-prop uses a top-down model to convert output error into plausible targets for weight updates.
result GAIT-prop and backpropagation give identical weight updates under certain conditions.
This work shows synthetic gradients can outperform backpropagation in sample efficiency.
problem The efficiency of backpropagation in training neural networks.
method Unified vectorized feedback framework for loss-based and reward-based learning, introducing synthetic gradients.
result Synthetic gradients can achieve lower gradient-estimation mean squared error than backpropagation under certain conditions.
A new method computes gradients without backpropagation.
problem Optimization of machine learning models.
method Forward mode automatic differentiation to compute gradients.
result Forward gradient is an unbiased estimate of the gradient, eliminating the need for backpropagation.
Improved autoencoder boosts sequence learning with less memory.
problem Training recurrent networks with high memory costs.
method Sparse Predictive Autoencoder (bRSM) with recurrent connections and boosting rule.
result Near optimal performance on stochastic sequence learning task.
Unified framework for faster neural network training with less information loss.
problem Time-consuming backpropagation and loss of unpropagated gradient information.
method Unified sparse backpropagation framework and memorized sparse backpropagation algorithm.
result Convergence in probability with certain conditions and effective information loss mitigation.
New algorithm shows neural networks can learn without full backpropagation.
problem Stochastic gradient descent with backpropagation is non-biologically plausible.
method Random and fixed backpropagation weights in a feedback alignment algorithm.
result Error converges to zero exponentially fast in overparameterized networks.