Interest in using biological neuronal networks as artificial learning systems has surged [e.g. 1,2,3]. However, networks of biological neurons differ from artificial neural networks: they are intrinsically noisy and have non-trivial physical embeddings. The question thus arises whether biological neurons are usable for artificial intelligence. A basic requirement for a system to be a good learning machine is the capacity to respond distinctly to different inputs, yet reproduce similar activity to similar stimuli. Liquid state [4] and reservoir [5] computing make direct and explicit use of this capacity by training a (linear) classifier on the input-responses of a system.In this study we use this strategy to investigate whether biological neuronal cultures have this response separation capacity. Merged and modular networks, resulting from an axon-growth algorithm, are simulated using the Izhikevich model. Inputs consist of hand written digits converted into spatio-temporal spike trains, and linear classifiers are trained to discriminate between two input classes (digits) from the culture's response. In addition, we confirm the obtained results in vitro using rat neurons cultured on micro-electrode arrays.The obtained results show that network activity is indeed different in response to different inputs and similar for similar inputs. In addition, we find that modular connectivity --characteristic of biological brains-- enhances the robustness of the performance of the network to distorted inputs and network alterations.[1] Kagan et al., Neuron (2022)[2] Sumi et al., PNAS (2023)[3] https://neuchip.eu[4] Maass et al., Neural Comput. (2002)[5] Jaeger, GMD Technical report (2001)