Alfredo Canziani

Since the emergence of Deep Neural Networks (DNNs) as a prominent technique in the field of computer vision, the ImageNet classification challenge has played a major role in advancing the state-of-the-art. While accuracy figures have steadily increased, the resource utilisation of winning models has not been properly taken into account. In this work, we present a comprehensive analysis of important metrics in practical applications: accuracy, memory footprint, parameters, operations count… Show more

Since the emergence of Deep Neural Networks (DNNs) as a prominent technique in the field of computer vision, the ImageNet classification challenge has played a major role in advancing the state-of-the-art. While accuracy figures have steadily increased, the resource utilisation of winning models has not been properly taken into account. In this work, we present a comprehensive analysis of important metrics in practical applications: accuracy, memory footprint, parameters, operations count, inference time and power consumption. Key findings are: (1) fully connected layers are largely inefficient for smaller batches of images; (2) accuracy and inference time are in a hyperbolic relationship; (3) energy constraint are an upper bound on the maximum achievable accuracy and model complexity; (4) the number of operations is a reliable estimate of the inference time. We believe our analysis provides a compelling set of information that helps design and engineer efficient DNNs. Show less

Animals use attentional mechanisms for being able to process enormous amount of sensory input in real time. Analogously, computerised systems could take advantage of similar techniques for achieving better timing performance. Visual attentional control uses bottom-up and top-down saliency maps for establishing the most relevant locations to observe. This article presents a novel fully-learnt unbiassed biologically plausible algorithm for computing both feature based and proto-object saliency… Show more

Animals use attentional mechanisms for being able to process enormous amount of sensory input in real time. Analogously, computerised systems could take advantage of similar techniques for achieving better timing performance. Visual attentional control uses bottom-up and top-down saliency maps for establishing the most relevant locations to observe. This article presents a novel fully-learnt unbiassed biologically plausible algorithm for computing both feature based and proto-object saliency maps, using a deep convolutional neural network simply trained on a single-class classification task, by unveiling its internal attentional apparatus. We are able to process 2 megapixels (MPs) colour images in real-time, i.e. at more than 10 frames per second, producing a 2MP map of interest. Show less

Oral presentation at Dawn or Doom 2014, Purdue University.

Dawn or Doom is a one-day seminar on the benefits and risks surrounding some of the technologies — such as artificial intelligence, nanotechnology, genetic engineering, and data science — which are both the most disruptive to current practices and being adopted the fastest.

Presentation abstract
What would it take to replicate the human visual system in synthetic hardware? What software models can we use to implement… Show more

Oral presentation at Dawn or Doom 2014, Purdue University.

Dawn or Doom is a one-day seminar on the benefits and risks surrounding some of the technologies — such as artificial intelligence, nanotechnology, genetic engineering, and data science — which are both the most disruptive to current practices and being adopted the fastest.

Presentation abstract
What would it take to replicate the human visual system in synthetic hardware? What software models can we use to implement the mammalian visual system? The goal of our research is a neuromorphic vision system capable of categorising, tracking and maintaining a visual memory of tens of targets. The application of such system is in smart phones, computers, robotics, autonomous cars, smart appliances, to name a few. Show less

For this study, monolithic piezoelectric sheets and macro fibre composite (MFC) generators were fixed to plates made of two materials commonly used for aircraft wing skin: Al-2024 aluminium alloy and an epoxy-carbon fibre composite. The plates then underwent harmonically varying loading in a tensile testing machine. The power generation of the harvesters was measured at a selection of strain levels and excitation frequencies, across a range of electrical loads. The optimal electrical load… Show more

For this study, monolithic piezoelectric sheets and macro fibre composite (MFC) generators were fixed to plates made of two materials commonly used for aircraft wing skin: Al-2024 aluminium alloy and an epoxy-carbon fibre composite. The plates then underwent harmonically varying loading in a tensile testing machine. The power generation of the harvesters was measured at a selection of strain levels and excitation frequencies, across a range of electrical loads. The optimal electrical load, yielding maximum power extraction, was identified for each working condition. The generated power increases quadratically with the strain and linearly with the frequency. The optimal electrical load decreases with increasing frequency and is only marginally dependent on strain. Show less