Research into the expressive power of neural networks typically considers real parameters and operations without rounding error. In this work, we study universal approximation property of quantized networks under discrete fixed-point parameters and fixed-point operations that may incur errors due to rounding. We first provide a necessary condition and a sufficient condition on fixed-point arithmetic and activation functions for universal approximation of quantized networks. Then, we show that various popular activation functions satisfy our sufficient condition, e.g., Sigmoid, ReLU, ELU, SoftPlus, SiLU, Mish, and GELU. In other words, networks using those activation functions are capable of universal approximation. We further show that our necessary condition and sufficient condition coincide under a mild condition on activation functions: e.g., for an activation function $\sigma$, there exists a fixed-point number $x$ such that $\sigma(x)=0$. Namely, we find a necessary and sufficient condition for a large class of activation functions. We lastly show that even quantized networks using binary weights in $\{-1,1\}$ can also universally approximate for practical activation functions.

When training neural networks with low-precision computation, rounding errors often cause stagnation or are detrimental to the convergence of the optimizers; in this paper we study the influence of rounding errors on the convergence of the gradient descent method for problems satisfying the Polyak-Lojasiewicz inequality. Within this context, we show that, in contrast, biased stochastic rounding errors may be beneficial since choosing a proper rounding strategy eliminates the vanishing gradient problem and forces the rounding bias in a descent direction. Furthermore, we obtain a bound on the convergence rate that is stricter than the one achieved by unbiased stochastic rounding. The theoretical analysis is validated by comparing the performances of various rounding strategies when optimizing several examples using low-precision fixed-point number formats.

Over the last few years, neural networks have started penetrating safety critical systems to take decisions in robots, rockets, autonomous driving car, etc. A problem is that these critical systems often have limited computing resources. Often, they use the fixed-point arithmetic for its many advantages (rapidity, compatibility with small memory devices.) In this article, a new technique is introduced to tune the formats (precision) of already trained neural networks using fixed-point arithmetic, which can be implemented using integer operations only. The new optimized neural network computes the output with fixed-point numbers without modifying the accuracy up to a threshold fixed by the user. A fixed-point code is synthesized for the new optimized neural network ensuring the respect of the threshold for any input vector belonging the range [xmin, xmax] determined during the analysis. From a technical point of view, we do a preliminary analysis of our floating neural network to determine the worst cases, then we generate a system of linear constraints among integer variables that we can solve by linear programming. The solution of this system is the new fixed-point format of each neuron. The experimental results obtained show the efficiency of our method which can ensure that the new fixed-point neural network has the same behavior as the initial floating-point neural network.

There is a recent interest in neural network (NN)-based communication algorithms which have shown to achieve (beyond) state-of-the-art performance for a variety of problems or lead to reduced implementation complexity. However, most work on this topic is simulation based and implementation on specialized hardware for fast inference, such as field-programmable gate arrays (FPGAs), is widely ignored. In particular for practical uses, NN weights should be quantized and inference carried out by a fixed-point instead of floating-point system, widely used in consumer class computers and graphics processing units (GPUs). Moving to such representations enables higher inference rates and complexity reductions, at the cost of precision loss. We demonstrate that it is possible to implement NN-based algorithms in fixed-point arithmetic with quantized weights at negligible performance loss and with hardware complexity compatible with practical systems, such as FPGAs and application-specific integrated circuits (ASICs).

The use of low-precision fixed-point arithmetic along with stochastic rounding has been proposed as a promising alternative to the commonly used 32-bit floating point arithmetic to enhance training neural networks training in terms of performance and energy efficiency. In the first part of this paper, the behaviour of the 12-bit fixed-point arithmetic when training a convolutional neural network with the CIFAR-10 dataset is analysed, showing that such arithmetic is not the most appropriate for the training phase. After that, the paper presents and evaluates, under the same conditions, alternative low-precision arithmetics, starting with the 12-bit floating-point arithmetic. These two representations are then leveraged using local scaling in order to increase accuracy and get closer to the baseline 32-bit floating-point arithmetic. Finally, the paper introduces a simplified model in which both the outputs and the gradients of the neural networks are constrained to power-of-two values, just using 7 bits for their representation. The evaluation demonstrates a minimal loss in accuracy for the proposed Power-of-Two neural network, avoiding the use of multiplications and divisions and thereby, significantly reducing the training time as well as the energy consumption and memory requirements during the training and inference phases.

Training of large-scale deep neural networks is often constrained by the available computational resources. We study the effect of limited precision data representation and computation on neural network training. Within the context of low-precision fixed-point computations, we observe the rounding scheme to play a crucial role in determining the network's behavior during training. Our results show that deep networks can be trained using only 16-bit wide fixed-point number representation when using stochastic rounding, and incur little to no degradation in the classification accuracy. We also demonstrate an energy-efficient hardware accelerator that implements low-precision fixed-point arithmetic with stochastic rounding.