Week 1: Computational Models of Reading and Acquired Dyslexia PDF

Week 1: Computational models of Reading and Acquired Dyslexia & Intro to Cognition Created time @January 17, 2025 1:39 PM Course Cognition and Emotion Type Lecture...

Week 1: Computational models of Reading and Acquired Dyslexia & Intro to Cognition Created time @January 17, 2025 1:39 PM Course Cognition and Emotion Type Lecture Archive Summary Key terms Questions Readings Lecture 1 (Dumay) Coltheart (1993) The DRC and PDP models were compared in this paper. Parallel-Distributed-Processing (PDP) Models: PDP models, also known as connectionist models/triangle models, suggest that reading aloud involves a network of interconnected processing units. These models learn to associate visual inputs (letters) with phonological outputs (sounds) through patterns of activation distributed across the network, rather than through explicit rules or lookup tables. Week 1: Computational models of Reading and Acquired Dyslexia & Intro to Cognition 1  No Explicit Rules: PDP models learn through experience and repeated exposure to written language. Distributed Representation: Information is represented by patterns of activation across many interconnected processing units, rather than a single, localised representation. Biological plausibility: PDP models align with our understanding of brain function since they emphasise distributed processing and learning through connections. The idea that cognitive processes are spread across neural networks is supported by neuroscience. Coltheart (2005) The non-connectionist DRC model was considered superior because it was more successful at simulating patterns of both acquired and developmental dyslexia. Explaining phenomena using DRC model… Regularity Effect: Irregular words are read slower than regular words because of conflict at the phoneme level between the two routes. Regular words do not generate conflict. Frequency Effects: High-frequency words are accessed faster in the mental lexicon, leading to faster reading aloud times Forster Chambers (1973) The study concludes that naming time is closely related to lexical decision time, suggesting that lexical search is involved in the naming process. The results indicate that the pronunciation of a word can be determined more rapidly by a dictionary look-up than by application of grapheme- phoneme rules, implying that lexical information can be retrieved before pronunciation. The study challenges the phonemic recoding hypothesis by demonstrating that lexical search does not require prior naming. Rayner (2010) Week 1: Computational models of Reading and Acquired Dyslexia & Intro to Cognition 2  Phonological dyslexia (difficulty with novel words/non-words): The DRC model attributes this to damage in the non-lexical route, while triangle models suggest damage to orthography-to-phonology connections. Surface dyslexia (difficulty with irregular words): The DRC model attributes this to damage in the lexical route, while triangle models attribute it to the orthography-to-phonology connections becoming overly specialised for pronouncing consistent words. https://www.canva.com/design/DAGcdyzqcHg/0l5YoOzJ-YP4IYeHZTtDeg/v iew REVISE: what regular/irregular word actually means? How to pronounce e.g. SEW using non-lexical route. rhymes with FEW. Definitions Reading is defined as information-processing: transforming print to speech, and/or at the same time, print to meaning. Both pathways are active all the time but depending on task one may be focused on. One goal of the science of reading is to uncover the components of this information-processing system. Computational models → the idea that by experimentation, we can design computer programs that simulate how we do it (i.e. how we read single words) Focus on single words for this module!! Week 1: Computational models of Reading and Acquired Dyslexia & Intro to Cognition 3  Acquired dyslexia, or alexia, is the partial or complete loss of the ability to read subsequent to a brain lesion. This is in contrast to developmental dyslexia, which is a lasting impairment in the acquisition of the ability to read. The Normal Reader There is a broad theoretical consensus that at least two different procedures accomplish the transformation from print to speech. In the beginning, it was recognised that we read in two ways: letter-by-letter for new words, and at a glance for common words (Saussure, 1922). How to Achieve pronunciation from written input: Not until 1970’s psychologists studied skill of reading. Forster & Chambers (1973) - two ways: 1. pronunciation computed by application of grapheme- phoneme rules. 2. pronunciation computed by searching LTM - direct dictionary lookup. The Dual-Route Cascaded (DRC) Model Week 1: Computational models of Reading and Acquired Dyslexia & Intro to Cognition 4  Non-lexical route Week 1: Computational models of Reading and Acquired Dyslexia & Intro to Cognition 5  The non-lexical route uses rules to relate segments of orthography (how a word is written) to segments of phonology (how it sounds). Essentially, it involves translating individual letters or groups of letters (graphemes) into their respective sounds and then assembling the sounds to build the word's pronunciation. This route does not involve the mental lexicon and can be used to pronounce words without recognising them. However, to understand the meaning, the word needs to be recognised. This route works with regular words and novel words, but not irregular words. The non-lexical route relies on grapheme-to-phoneme correspondence (GPC) rules to derive pronunciation. For instance, RIGHT = R-IGH-T = /r-aI-t/. You are actively COMPUTING words rather than retrieving from memory. Through knowledge of english we know how groups of letters sound e.g. IGH = I. There is rules but there are also exceptions you have to learn. This non-lexical route would be good for many words except exceptions. E.g. HAVE vs CAVE This will lead to error. But allows for approximation of novel word pronounciation! Lexical Route Week 1: Computational models of Reading and Acquired Dyslexia & Intro to Cognition 6  Orthographic input lexicon = memory of how words are spelt. Week 1: Computational models of Reading and Acquired Dyslexia & Intro to Cognition 7  Phonological output lexicon = memory of how words sound. The lexical route looks up words in long-term memory to retrieve knowledge about their meaning and pronunciation. It relies on visual word recognition to access the mental lexicon. This route works for regular and irregular words, but not for novel or unknown words because they are not in the reader's mental dictionary. Both the lexical and non-lexical routes make use of excitatory and inhibitory connections across levels of representations. Excitatory connections activate compatible representations at the next level and sometimes at the preceding level. For example, seeing the letter "P" activates all words in the input lexicon that contain "P", like "PEN", and suppresses those that do not, like "STOOL". Inhibitory connections suppress representations that are not compatible. Feedback influence occurs when a word in the input lexicon, such as "PEN", further activates the letter units "P", "E", and "N" and suppresses all other letters. Whether the connection is excitatory or inhibitory (e.g. reaching letter units) depends on before component (e.g. what visual feature units). Excitatory when the stimuli corresponds to what you see on the page. Inhibitory is turning down response for other features not present. So, we see the p stimulus. Then correspond with matching letter. Then excitatory connections to activate words in the input lexicon that contain ‘p’ (PEN), inhibitory to words not containing ‘p’ → stool. These are ‘feedforward connections’ because they are going down. Feedback influences → reverse as well. Excitatory to other letters including E, N that may be the next letters. Inhibitory to other letters (S, O etc.) Week 1: Computational models of Reading and Acquired Dyslexia & Intro to Cognition 8  Retroactive feedback. Shows memory can influence perception. In the non-lexical route, their is no feedback influences past letter units. The model does not decide which route is used; the output is produced by the model as a whole. You activate both routes at the same time. Non-words Even for non-words (e.g. SARE) the lexical route will activate meanings for similar words like CARE, SORE Nonwords, or made-up words, do not have an entry in long-term memory. However, they may still evoke existing words due to visual similarities, leading to false recognition. Their proximity to an exception word may also influence pronunciation. For example, FINT is more likely to be read as PINT (exception) than as MINT (regular pronunciation), because FINT is more visually similar to PINT than it is to MINT. Pseudoword reading allows for the assessment of knowledge about print-to-sound correspondences. Time pressure The two routes do not race against each other except when reading under time pressure. Under time pressure, you may regularize an exception word or mistake a nonword for a word. The both routes aren’t racing, unless reading under time pressure. Week 1: Computational models of Reading and Acquired Dyslexia & Intro to Cognition 9  When under time pressure, errors can be made due to this lexical route activating meanings for similar words. Language Transparency The transparency/opacity of a language affects the demands on the lexical vs. nonlexical route. Transparent languages (e.g., Dutch, Spanish, Italian) have consistent letter-sound correspondences, making it easier for children to learn to read. Opaque languages (e.g., English, French, Danish, Portuguese) have less consistent letter-sound correspondences and a higher prevalence of developmental dyslexia. Week 1: Computational models of Reading and Acquired Dyslexia & Intro to Cognition 10  Key effects found in both DRC and human readers Words are read faster than non-words. Why? Activation of lexical route means that they have the whole system to help identify word. Non-words only have half the system (non-lexical) High-frequency words are read faster than low-frequency words. Why? More accessible from lexicon due to practice effects. Regular words are read faster and more accurately than irregular words, especially for low-frequency words. Why? As the irregular words can only be read really via lexical route. Regular can use Graphene-Phoneme correspondence rules and lexical route. Low frequency is less as not practiced often - exception is not recognised as well. Non-words with larger orthographic neighbourhoods (i.e., similar to many words) are read faster. Why? E.g. Lat vs Hom → Lat has many neighbours like rat, cat, fat. So orthographic knowledge from this helps to get meaning. Hom doesn’t have many neighbours. Non-words that sound like words are read faster than those that do not. Why? E.g. Brane/grune - even though brane isn’t stored in lexicon, the non-lexical route allows to gather phonology which then is used to find meaning. Lexicon can then help reinforce pronunciation. Longer non-words take longer to read, but the number of letters has little to no effect on reading real words. Week 1: Computational models of Reading and Acquired Dyslexia & Intro to Cognition 11  Why? Non-lexical route is a computational route and has to work harder for more letters so it takes longer. These models have results from 3 strands of research: 1. Patients who have disorders of reading (developmental and acquired dyslexia) 2. How normal readers function in tasks 3. Computer simulations on how humans deal with written info. Summary Irregular words e.g. HAVE, can only be read correctly using whole- word, long-term knowledge. (Lexical Route) Nonwords e.g. PLICK, (or unknown words) are read using only knowledge of print-to-sound correspondences, unless they are very similar to a word. (Non-lexical route) Regular words e.g. CAVE, can be read by both routes simultaneously without conflict. (Lexical and non-lexical route) https://www.canva.com/design/DAGcd3RMBp8/jHv9jXwBEr8JFFenkVrR4 w/view Week 1: Computational models of Reading and Acquired Dyslexia & Intro to Cognition 12  A (Very) Brief History of Psychology Behaviourism (1950s-60s) This school of thought, championed by figures like Skinner and Watson, posited that psychology should solely focus on observable behaviors. They believed that behavior could be explained without considering internal mental events. The emphasis was on understanding the relationship between inputs (external stimuli) and outputs (behavior), essentially attempting to define the rules of learning. For example, language development was seen as a product of learned responses to stimuli. Challenges to Behaviourism: language development couldn't be adequately explained through learned responses alone. Others argued that the inherent complexity of language acquisition necessitates internal mental processes. The emergence of computers offered a new analogy for understanding the mind, paving the way for cognitive science. Cognitive Science (1970s/80s) This era saw the mind being conceptualized as software operating on the hardware of the brain. Cognitive scientists, while still studying behavior through measures like reaction times, started inferring the existence of "mental modules" responsible for specific cognitive functions. Brain as information processing machine/computer. Cognitive Neuropsychology (1980s/90s) Advances in imaging techniques like MRI, coupled with observations of specific functional losses following brain lesions, led to the rise of cognitive neuropsychology. This field emphasises the localisation of internal mental events to specific brain regions. Classic examples include Broca's area, associated with language production, and Wernicke's area, linked to language comprehension. The concept of double dissociation, where lesions in different areas lead to distinct and opposite patterns of impairment, strengthened the idea of functional localisation. Cognitive Neuroscience (1990s onwards) Week 1: Computational models of Reading and Acquired Dyslexia & Intro to Cognition 13  Driven by tools for measuring and manipulating brain function, cognitive neuroscience integrated the localizationist perspective of neuropsychology with the precision of techniques like fMRI and TMS. A central aim of this field is to uncover the neural underpinnings of behavior – the search for the mind within the brain. Methods of Cognitive Neuroscience Depends on converging evidence from different methods. Activation Methods: These techniques, such as functional Magnetic Resonance Imaging (fMRI) and Electroencephalography (EEG), allow researchers to observe brain activity during cognitive tasks. fMRI offers high spatial resolution (ability to pinpoint activity) but has lower temporal resolution (ability to track changes over time). EEG, on the other hand, has excellent temporal resolution but suffers from poor spatial resolution. Importantly, activation methods primarily reveal correlations between brain activity and behavior, not causation. But we can ask other important questions like how does a process actually work? Deactivation Methods: Transcranial Magnetic Stimulation (TMS) and neuropsychology (lesion-deficit mapping) fall under this category, allowing for inferences about the necessity of specific brain regions for particular cognitive functions. TMS creates "virtual lesions" by disrupting neural activity in targeted areas, offering high temporal precision. However, its reach is limited to cortical regions near the scalp. Neuropsychology studies individuals with brain damage, often relying on double dissociation logic or lesion mapping techniques with larger samples. While providing causal evidence, the spatial precision of neuropsychological findings can be limited due to the size of lesioned areas. Monkey single unit (neuron) recordings - Monkey is trained on computerised task. Put to sleep and electrode array inserted into brain. When awoken, monkey performs task while neuronal activity recorded. Pro - temporal and spatial resolution. Con - unethical and translation to humans. Key Questions in Cognitive Neuroscience Week 1: Computational models of Reading and Acquired Dyslexia & Intro to Cognition 14  Can we reduce all mental phenomena to brain processes? Cognitive neuroscience doesn’t require you to accept that complex mental phenomena can be reduced to firing of single neurons - but theres an intermediate level → neuronal systems. They don’t have to be mapped to brain regions (but can be) Functional Localisation: A primary focus has been mapping cognitive processes onto brain regions. It's a two-step process: 1. Defining the cognitive ontology (basic building blocks of cognition) 2. mapping it onto brain regions Keep going between both steps. Mechanics: Cognitive neuroscience seeks to understand how different brain areas mediate cognitive functions. This involves investigating the mechanisms underlying cognitive processes, like the limitations of working memory capacity. Connectivity: How different brain modules interact and communicate with each other. Understanding these connections is essential for unraveling the complexity of the brain. Course Overview This course will delve into a range of cognitive domains, including: Object Recognition: Exploring how we recognise visual objects, including potential pathways and disorders like agnosia. The special case of face recognition will also be discussed. Attention: Understanding how we selectively process relevant information while filtering out irrelevant stimuli, including mechanisms and breakdowns of attention like hemispatial neglect. Working Memory: Investigating the capacity to maintain information over short periods, its relationship to intelligence, and the role of the prefrontal cortex. Week 1: Computational models of Reading and Acquired Dyslexia & Intro to Cognition 15  Prefrontal Cortex and Executive Functions: Examining the control of actions, inhibition of inappropriate responses, and the concept of an "inhibitory control" module. Emotion: Exploring models of emotion generation, the role of the amygdala, and the interplay of emotion with learning, attention, and decision making. Relationship Between Basic Science and Clinical Implications The course will highlight the bidirectional relationship between basic cognitive neuroscience research and clinical applications. Clinical disorders offer insights into the workings of the healthy brain, while basic science findings can shed light on clinical conditions. For example, studying agnosia informs our understanding of object recognition, while knowledge of attentional mechanisms can aid in comprehending hemispatial neglect. Tools of Cognitive Neuroscience Beyond the brain imaging and lesion studies mentioned earlier, the course will also utilise computerised cognitive testing. This involves measuring reaction times and/or accuracy across different experimental conditions to infer the operation of specific cognitive processes. The subtraction logic posits that the difference in performance between two conditions differing by a single process reflects the contribution of that process. For instance, comparing reaction times in a task requiring selective attention to a control condition can isolate the effect of attention. Critical analysis key tip → By subtraction logic - some people may say that the difference in performance reflects difference in process. But others may say does it? other processes? i.e. how well are we controlling for confound variables. Stroke Fatty deposits restrict carotid artery - restricts blood flow to brain. Blood pressure increases. Week 1: Computational models of Reading and Acquired Dyslexia & Intro to Cognition 16  High blood pressure pushes Thrombus that is lodged in cerebral artery. Restrict blood flow to part of brain. fMRI Measures Blood Oxygen Level Dependant Signal - a proxy for neuronal activity. Measuring magnetic properties of oxygenated blood. Differences in these magnetic properties → not actually direct neural activity. High spatial resolution (mm3), low temporal (seconds). Key Points to Remember Cognitive neuroscience emerged from the evolution of psychological thought, building upon behaviorism and cognitive science. The field employs a diverse toolkit, including activation and deactivation methods, each with strengths and limitations. Cognitive neuroscience addresses key questions about functional localisation, the mechanics of cognitive processes, and brain connectivity. This course will explore core cognitive domains, emphasising the interplay between basic science and clinical insights. Computerised cognitive testing, along with other tools, will be used to investigate brain-behavior relationships. Week 1: Computational models of Reading and Acquired Dyslexia & Intro to Cognition 17