id	author	title	date	pages	extension	mime	words	sentences	flesch	summary	cache	txt
cord-031663-i71w0es7	Giacobbe, Mirco	How Many Bits Does it Take to Quantize Your Neural Network?	2020-03-13		.txt	text/plain	6525	332	53	For this reason, we introduce a verification method for quantized neural networks which, using SMT solving over bit-vectors, accounts for their exact, bit-precise semantics. As a result, we obtain a encoding into a first-order logic formula which, in contrast to a standard unbalanced linear encoding, makes the verification of quantized networks practical and amenable to modern bit-precise SMT-solving. We measured the robustness to attacks of a neural classifier involving 890 neurons and trained on the MNIST dataset (handwritten digits), for quantizations between 6 and 10 bits. We evaluated whether our balanced encoding strategy, compared to a standard linear encoding, can improve the scalability of contemporary SMT solvers for quantifier-free bit-vectors (QF BV) to check specifications of quantized neural networks. We introduced the first complete method for the verification of quantized neural networks which, by SMT solving over bit-vectors, accounts for their bit-precise semantics.	./cache/cord-031663-i71w0es7.txt	./txt/cord-031663-i71w0es7.txt