| Description |
Please join us at the next MMM – Werrington South edition with invited speaker Professor Paul Hurley as he presents on the ‘Imaging of Things’, followed by an exciting project presentation by our ICNS-IISc Interns as their program draws to an end.Venue: Level 2, Building BA, (Large meeting space), Werrington South Campus Topic: Imaging of Things
Speaker: Professor Paul Hurley, Professor in Data Science, School of Computing, Engineering and Mathematics
Abstract: We, and cognitive systems in general, learn and adapt to our surroundings. Experience determines what to sense. Focus is on what we care about, forgoing all deemed irrelevant (no one likes information overload). For example, the visual cortex combines a blurry image with prior knowledge to instruct the eye on what to look at next, so as to extract further information accurately using as little energy as possible.
In this short talk, I will introduce how, by the right level of abstraction, we can build efficient, adaptive algorithms that are holisitic: taking the sensor and the end knowledge extraction into consideration. The framework has applications across the spectrum of radio astronomy, audio, medical and seismic imaging, internet-of-things etc.
Topic: Auditory Event-based Unsupervised Feature Extraction
Speakers: Ying Xu, Akshaya A Mukundan, Shilpa Kontham Kulangara
Abstract: In human auditory pathway, acoustic information is extracted and conveyed through sequence of action potentials, or spikes. The spike streams form robust representations for encoding auditory features that are important for recognition. For machine hearing related applications, it is necessary to extract the acoustic features of the audio source. Spike-based (event-based) learning algorithms are highly efficient in detection of such spatio-temporal information, compared to other conventional frame-based architectures. Each spike event can be represented by its unique features and firing time, resulting in significant data compression. To investigate and understand signal processing in the human auditory system, binaural Cascade of Asymmetric Resonators with Fast-Acting Compression (CAR-FAC) cochlear system combined with stochastic leaky integrate and fire (LIF) neurons is used to generate spike streams from sound input. Further, we examine and investigate existing and novel spike processing methods for processing auditory spike streams. We used the event-based unsupervised feature extraction approach that has been used in the event-based vision processing to extract acoustic characteristics.
Topic: Auditory Attention model based on Saliency maps
Speakers: Ying Xu, Muhammed Hafees Hak, Siddharth Chaudhary
Abstract: It is fascinating to note that the biological system uses saliency-based attention to detect in real time, general conspicuous targets from cluttered scenes and process the relevant information for further making decisions. Incorporating such detection and information processing capability in artificial systems has important applications such as sound segmentation. We propose a biological inspired hardware model of auditory attention, captured by auditory saliency map. It uses the cochlear model, CAR-FAC, for the early auditory processing and a cortical model for extracting the spectro-temporal features from the audio input. These features are further inhibited to promote the salient information and further combined to generate the auditory saliency map. The novelty of the approach lies in the inhibition algorithm to suppress the non-salient information. This work further explores the possibility of FPGA implementation of the design.
IMPORTANT
Please note, there will be no MMM or MAC on 5 August, 2019. We thank you for your continued attendance at the MMM's. Let's keep the momentum going. MARCS staff and students are reminded that all meetings and workshops have an important role in building and maintaining the sense of community which is central to the success of MARCS as a cooperative and energetic research institute. Your attendance is both welcomed and expected. The zoom ID is: 627 146 998. Link to zoom https://uws.zoom.us/j/627146998 |
